Pipeline cache API and implementation for Vulkan (#5319)

Co-authored-by: Connor Fitzgerald <connorwadefitzgerald@gmail.com>

CHANGELOG.md

@@ -72,6 +72,10 @@ By @stefnotch in [#5410](https://github.com/gfx-rs/wgpu/pull/5410)
 
 ### New features
 
+#### Vulkan
+
+- Added a `PipelineCache` resource to allow using Vulkan pipeline caches. By @DJMcNab in [#5319](https://github.com/gfx-rs/wgpu/pull/5319)
+
 #### General
 
 #### Naga

benches/benches/renderpass.rs

@@ -207,6 +207,7 @@ impl RenderpassState {
                         compilation_options: wgpu::PipelineCompilationOptions::default(),
                     }),
                     multiview: None,
+                    cache: None,
                 });
 
         let render_target = device_state
@@ -304,6 +305,7 @@ impl RenderpassState {
                         compilation_options: wgpu::PipelineCompilationOptions::default(),
                     }),
                     multiview: None,
+                    cache: None,
                 },
             ));
         }

deno_webgpu/pipeline.rs

@@ -115,6 +115,7 @@ pub fn op_webgpu_create_compute_pipeline(
             constants: Cow::Owned(compute.constants.unwrap_or_default()),
             zero_initialize_workgroup_memory: true,
         },
+        cache: None,
     };
     let implicit_pipelines = match layout {
         GPUPipelineLayoutOrGPUAutoLayoutMode::Layout(_) => None,
@@ -395,6 +396,7 @@ pub fn op_webgpu_create_render_pipeline(
         multisample: args.multisample,
         fragment,
         multiview: None,
+        cache: None,
     };
 
     let implicit_pipelines = match args.layout {

examples/src/boids/mod.rs

@@ -156,6 +156,7 @@ impl crate::framework::Example for Example {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         // create compute pipeline
@@ -166,6 +167,7 @@ impl crate::framework::Example for Example {
             module: &compute_shader,
             entry_point: "main",
             compilation_options: Default::default(),
+            cache: None,
         });
 
         // buffer for the three 2d triangle vertices of each instance

examples/src/bunnymark/mod.rs

@@ -224,6 +224,7 @@ impl crate::framework::Example for Example {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         let texture = {

examples/src/conservative_raster/mod.rs

@@ -113,6 +113,7 @@ impl crate::framework::Example for Example {
                 depth_stencil: None,
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
 
         let pipeline_triangle_regular =
@@ -135,6 +136,7 @@ impl crate::framework::Example for Example {
                 depth_stencil: None,
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
 
         let pipeline_lines = if device
@@ -165,6 +167,7 @@ impl crate::framework::Example for Example {
                     depth_stencil: None,
                     multisample: wgpu::MultisampleState::default(),
                     multiview: None,
+                    cache: None,
                 }),
             )
         } else {
@@ -224,6 +227,7 @@ impl crate::framework::Example for Example {
                     depth_stencil: None,
                     multisample: wgpu::MultisampleState::default(),
                     multiview: None,
+                    cache: None,
                 }),
                 bind_group_layout,
             )

examples/src/cube/mod.rs

@@ -260,6 +260,7 @@ impl crate::framework::Example for Example {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         let pipeline_wire = if device
@@ -301,6 +302,7 @@ impl crate::framework::Example for Example {
                 depth_stencil: None,
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
             Some(pipeline_wire)
         } else {

examples/src/hello_compute/mod.rs

@@ -110,6 +110,7 @@ async fn execute_gpu_inner(
         module: &cs_module,
         entry_point: "main",
         compilation_options: Default::default(),
+        cache: None,
     });
 
     // Instantiates the bind group, once again specifying the binding of buffers.

examples/src/hello_synchronization/mod.rs

@@ -104,6 +104,7 @@ async fn execute(
         module: &shaders_module,
         entry_point: "patient_main",
         compilation_options: Default::default(),
+        cache: None,
     });
     let hasty_pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
         label: None,
@@ -111,6 +112,7 @@ async fn execute(
         module: &shaders_module,
         entry_point: "hasty_main",
         compilation_options: Default::default(),
+        cache: None,
     });
 
     //----------------------------------------------------------

examples/src/hello_triangle/mod.rs

@@ -72,6 +72,7 @@ async fn run(event_loop: EventLoop<()>, window: Window) {
         depth_stencil: None,
         multisample: wgpu::MultisampleState::default(),
         multiview: None,
+        cache: None,
     });
 
     let mut config = surface

examples/src/hello_workgroups/mod.rs

@@ -111,6 +111,7 @@ async fn run() {
         module: &shader,
         entry_point: "main",
         compilation_options: Default::default(),
+        cache: None,
     });
 
     //----------------------------------------------------------

examples/src/mipmap/mod.rs

@@ -109,6 +109,7 @@ impl Example {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         let bind_group_layout = pipeline.get_bind_group_layout(0);
@@ -310,6 +311,7 @@ impl crate::framework::Example for Example {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         // Create bind group

examples/src/msaa_line/mod.rs

@@ -78,6 +78,7 @@ impl Example {
                 ..Default::default()
             },
             multiview: None,
+            cache: None,
         });
         let mut encoder =
             device.create_render_bundle_encoder(&wgpu::RenderBundleEncoderDescriptor {

examples/src/render_to_texture/mod.rs

@@ -72,6 +72,7 @@ async fn run(_path: Option<String>) {
         depth_stencil: None,
         multisample: wgpu::MultisampleState::default(),
         multiview: None,
+        cache: None,
     });
 
     log::info!("Wgpu context set up.");

examples/src/repeated_compute/mod.rs

@@ -246,6 +246,7 @@ impl WgpuContext {
             module: &shader,
             entry_point: "main",
             compilation_options: Default::default(),
+            cache: None,
         });
 
         WgpuContext {

examples/src/shadow/mod.rs

@@ -526,6 +526,7 @@ impl crate::framework::Example for Example {
                 }),
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
 
             Pass {
@@ -660,6 +661,7 @@ impl crate::framework::Example for Example {
                 }),
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
 
             Pass {

examples/src/skybox/mod.rs

@@ -221,6 +221,7 @@ impl crate::framework::Example for Example {
             }),
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
         let entity_pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
             label: Some("Entity"),
@@ -254,6 +255,7 @@ impl crate::framework::Example for Example {
             }),
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         let sampler = device.create_sampler(&wgpu::SamplerDescriptor {

examples/src/srgb_blend/mod.rs

@@ -151,6 +151,7 @@ impl<const SRGB: bool> crate::framework::Example for Example<SRGB> {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         // Done

examples/src/stencil_triangles/mod.rs

@@ -106,6 +106,7 @@ impl crate::framework::Example for Example {
             }),
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         let outer_pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
@@ -141,6 +142,7 @@ impl crate::framework::Example for Example {
             }),
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
         let stencil_buffer = device.create_texture(&wgpu::TextureDescriptor {

examples/src/storage_texture/mod.rs

@@ -101,6 +101,7 @@ async fn run(_path: Option<String>) {
         module: &shader,
         entry_point: "main",
         compilation_options: Default::default(),
+        cache: None,
     });
 
     log::info!("Wgpu context set up.");

examples/src/texture_arrays/mod.rs

@@ -341,6 +341,7 @@ impl crate::framework::Example for Example {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None
         });
 
         Self {

examples/src/timestamp_queries/mod.rs

@@ -299,6 +299,7 @@ fn compute_pass(
         module,
         entry_point: "main_cs",
         compilation_options: Default::default(),
+        cache: None,
     });
     let bind_group_layout = compute_pipeline.get_bind_group_layout(0);
     let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
@@ -366,8 +367,8 @@ fn render_pass(
         depth_stencil: None,
         multisample: wgpu::MultisampleState::default(),
         multiview: None,
+        cache: None,
     });
-
     let render_target = device.create_texture(&wgpu::TextureDescriptor {
         label: Some("rendertarget"),
         size: wgpu::Extent3d {

examples/src/uniform_values/mod.rs

@@ -192,8 +192,8 @@ impl WgpuContext {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
-
         let surface_config = surface
             .get_default_config(&adapter, size.width, size.height)
             .unwrap();

examples/src/water/mod.rs

@@ -574,6 +574,8 @@ impl crate::framework::Example for Example {
             // No multisampling is used.
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            // No pipeline caching is used
+            cache: None,
         });
 
         // Same idea as the water pipeline.
@@ -610,6 +612,7 @@ impl crate::framework::Example for Example {
             }),
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None
         });
 
         // A render bundle to draw the terrain.

player/src/lib.rs

@@ -302,6 +302,12 @@ impl GlobalPlay for wgc::global::Global {
             Action::DestroyRenderPipeline(id) => {
                 self.render_pipeline_drop::<A>(id);
             }
+            Action::CreatePipelineCache { id, desc } => {
+                let _ = unsafe { self.device_create_pipeline_cache::<A>(device, &desc, Some(id)) };
+            }
+            Action::DestroyPipelineCache(id) => {
+                self.pipeline_cache_drop::<A>(id);
+            }
             Action::CreateRenderBundle { id, desc, base } => {
                 let bundle =
                     wgc::command::RenderBundleEncoder::new(&desc, device, Some(base)).unwrap();

tests/src/image.rs

@@ -370,6 +370,7 @@ fn copy_via_compute(
         module: &sm,
         entry_point: "copy_texture_to_buffer",
         compilation_options: Default::default(),
+        cache: None,
     });
 
     {

tests/tests/bgra8unorm_storage.rs

@@ -98,6 +98,7 @@ static BGRA8_UNORM_STORAGE: GpuTestConfiguration = GpuTestConfiguration::new()
             entry_point: "main",
             compilation_options: Default::default(),
             module: &module,
+            cache: None,
         });
 
         let mut encoder =

tests/tests/bind_group_layout_dedup.rs

@@ -91,6 +91,7 @@ async fn bgl_dedupe(ctx: TestingContext) {
             module: &module,
             entry_point: "no_resources",
             compilation_options: Default::default(),
+            cache: None,
         };
 
         let pipeline = ctx.device.create_compute_pipeline(&desc);
@@ -220,6 +221,7 @@ fn bgl_dedupe_with_dropped_user_handle(ctx: TestingContext) {
             module: &module,
             entry_point: "no_resources",
             compilation_options: Default::default(),
+            cache: None,
         });
 
     let mut encoder = ctx.device.create_command_encoder(&Default::default());
@@ -266,6 +268,7 @@ fn bgl_dedupe_derived(ctx: TestingContext) {
             module: &module,
             entry_point: "resources",
             compilation_options: Default::default(),
+            cache: None,
         });
 
     // We create two bind groups, pulling the bind_group_layout from the pipeline each time.
@@ -337,6 +340,7 @@ fn separate_programs_have_incompatible_derived_bgls(ctx: TestingContext) {
         module: &module,
         entry_point: "resources",
         compilation_options: Default::default(),
+        cache: None,
     };
     // Create two pipelines, creating a BG from the second.
     let pipeline1 = ctx.device.create_compute_pipeline(&desc);
@@ -399,6 +403,7 @@ fn derived_bgls_incompatible_with_regular_bgls(ctx: TestingContext) {
             module: &module,
             entry_point: "resources",
             compilation_options: Default::default(),
+            cache: None,
         });
 
     // Create a matching BGL

tests/tests/buffer.rs

@@ -225,6 +225,7 @@ static MINIMUM_BUFFER_BINDING_SIZE_LAYOUT: GpuTestConfiguration = GpuTestConfigu
                     module: &shader_module,
                     entry_point: "main",
                     compilation_options: Default::default(),
+                    cache: None,
                 });
         });
     });
@@ -294,6 +295,7 @@ static MINIMUM_BUFFER_BINDING_SIZE_DISPATCH: GpuTestConfiguration = GpuTestConfi
                 module: &shader_module,
                 entry_point: "main",
                 compilation_options: Default::default(),
+                cache: None,
             });
 
         let buffer = ctx.device.create_buffer(&wgpu::BufferDescriptor {

tests/tests/compute_pass_resource_ownership.rs

@@ -161,6 +161,7 @@ fn resource_setup(ctx: &TestingContext) -> ResourceSetup {
             module: &sm,
             entry_point: "main",
             compilation_options: Default::default(),
+            cache: None,
         });
 
     ResourceSetup {

tests/tests/device.rs

@@ -488,6 +488,7 @@ static DEVICE_DESTROY_THEN_MORE: GpuTestConfiguration = GpuTestConfiguration::ne
                     multisample: wgpu::MultisampleState::default(),
                     fragment: None,
                     multiview: None,
+                    cache: None,
                 });
         });
 
@@ -500,6 +501,7 @@ static DEVICE_DESTROY_THEN_MORE: GpuTestConfiguration = GpuTestConfiguration::ne
                     module: &shader_module,
                     entry_point: "",
                     compilation_options: Default::default(),
+                    cache: None,
                 });
         });
 
@@ -757,6 +759,7 @@ fn vs_main() -> @builtin(position) vec4<f32> {
                 depth_stencil: None,
                 multisample: wgt::MultisampleState::default(),
                 multiview: None,
+                cache: None
             });
 
         // fail(&ctx.device, || {

tests/tests/mem_leaks.rs

@@ -113,6 +113,7 @@ async fn draw_test_with_reports(
                 })],
             }),
             multiview: None,
+            cache: None,
         });
 
     let global_report = ctx.instance.generate_report().unwrap();

tests/tests/nv12_texture/mod.rs

@@ -41,6 +41,7 @@ static NV12_TEXTURE_CREATION_SAMPLING: GpuTestConfiguration = GpuTestConfigurati
                 depth_stencil: None,
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
 
         let tex = ctx.device.create_texture(&wgpu::TextureDescriptor {

tests/tests/occlusion_query/mod.rs

@@ -51,6 +51,7 @@ static OCCLUSION_QUERY: GpuTestConfiguration = GpuTestConfiguration::new()
                 }),
                 multisample: wgpu::MultisampleState::default(),
                 multiview: None,
+                cache: None,
             });
 
         // Create occlusion query set

tests/tests/partially_bounded_arrays/mod.rs

@@ -70,6 +70,7 @@ static PARTIALLY_BOUNDED_ARRAY: GpuTestConfiguration = GpuTestConfiguration::new
             module: &cs_module,
             entry_point: "main",
             compilation_options: Default::default(),
+            cache: None,
         });
 
         let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {

tests/tests/pipeline.rs

@@ -29,6 +29,7 @@ static PIPELINE_DEFAULT_LAYOUT_BAD_MODULE: GpuTestConfiguration = GpuTestConfigu
                     module: &module,
                     entry_point: "doesn't exist",
                     compilation_options: Default::default(),
+                    cache: None,
                 });
 
             pipeline.get_bind_group_layout(0);

tests/tests/pipeline_cache.rs

@@ -0,0 +1,192 @@
+use std::{fmt::Write, num::NonZeroU64};
+
+use wgpu_test::{gpu_test, GpuTestConfiguration, TestParameters, TestingContext};
+
+/// We want to test that using a pipeline cache doesn't cause failure
+///
+/// It would be nice if we could also assert that reusing a pipeline cache would make compilation
+/// be faster however, some drivers use a fallback pipeline cache, which makes this inconsistent
+/// (both intra- and inter-run).
+#[gpu_test]
+static PIPELINE_CACHE: GpuTestConfiguration = GpuTestConfiguration::new()
+    .parameters(
+        TestParameters::default()
+            .test_features_limits()
+            .features(wgpu::Features::PIPELINE_CACHE),
+    )
+    .run_async(pipeline_cache_test);
+
+/// Set to a higher value if adding a timing based assertion. This is otherwise fast to compile
+const ARRAY_SIZE: u64 = 256;
+
+/// Create a shader which should be slow-ish to compile
+fn shader() -> String {
+    let mut body = String::new();
+    for idx in 0..ARRAY_SIZE {
+        // "Safety": There will only be a single workgroup, and a single thread in that workgroup
+        writeln!(body, "    output[{idx}] = {idx}u;")
+            .expect("`u64::fmt` and `String::write_fmt` are infallible");
+    }
+
+    format!(
+        r#"
+        @group(0) @binding(0)
+        var<storage, read_write> output: array<u32>;
+    
+        @compute @workgroup_size(1)
+        fn main() {{
+        {body}
+        }}
+        "#,
+    )
+}
+
+async fn pipeline_cache_test(ctx: TestingContext) {
+    let shader = shader();
+    let sm = ctx
+        .device
+        .create_shader_module(wgpu::ShaderModuleDescriptor {
+            label: Some("shader"),
+            source: wgpu::ShaderSource::Wgsl(shader.into()),
+        });
+
+    let bgl = ctx
+        .device
+        .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+            label: Some("bind_group_layout"),
+            entries: &[wgpu::BindGroupLayoutEntry {
+                binding: 0,
+                visibility: wgpu::ShaderStages::COMPUTE,
+                ty: wgpu::BindingType::Buffer {
+                    ty: wgpu::BufferBindingType::Storage { read_only: false },
+                    has_dynamic_offset: false,
+                    min_binding_size: NonZeroU64::new(ARRAY_SIZE * 4),
+                },
+                count: None,
+            }],
+        });
+
+    let gpu_buffer = ctx.device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some("gpu_buffer"),
+        size: ARRAY_SIZE * 4,
+        usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
+        mapped_at_creation: false,
+    });
+
+    let cpu_buffer = ctx.device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some("cpu_buffer"),
+        size: ARRAY_SIZE * 4,
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
+        mapped_at_creation: false,
+    });
+
+    let bind_group = ctx.device.create_bind_group(&wgpu::BindGroupDescriptor {
+        label: Some("bind_group"),
+        layout: &bgl,
+        entries: &[wgpu::BindGroupEntry {
+            binding: 0,
+            resource: gpu_buffer.as_entire_binding(),
+        }],
+    });
+
+    let pipeline_layout = ctx
+        .device
+        .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+            label: Some("pipeline_layout"),
+            bind_group_layouts: &[&bgl],
+            push_constant_ranges: &[],
+        });
+
+    let first_cache_data;
+    {
+        let first_cache = unsafe {
+            ctx.device
+                .create_pipeline_cache(&wgpu::PipelineCacheDescriptor {
+                    label: Some("pipeline_cache"),
+                    data: None,
+                    fallback: false,
+                })
+        };
+        let first_pipeline = ctx
+            .device
+            .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
+                label: Some("pipeline"),
+                layout: Some(&pipeline_layout),
+                module: &sm,
+                entry_point: "main",
+                compilation_options: Default::default(),
+                cache: Some(&first_cache),
+            });
+        validate_pipeline(&ctx, first_pipeline, &bind_group, &gpu_buffer, &cpu_buffer).await;
+        first_cache_data = first_cache.get_data();
+    }
+    assert!(first_cache_data.is_some());
+
+    let second_cache = unsafe {
+        ctx.device
+            .create_pipeline_cache(&wgpu::PipelineCacheDescriptor {
+                label: Some("pipeline_cache"),
+                data: first_cache_data.as_deref(),
+                fallback: false,
+            })
+    };
+    let first_pipeline = ctx
+        .device
+        .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
+            label: Some("pipeline"),
+            layout: Some(&pipeline_layout),
+            module: &sm,
+            entry_point: "main",
+            compilation_options: Default::default(),
+            cache: Some(&second_cache),
+        });
+    validate_pipeline(&ctx, first_pipeline, &bind_group, &gpu_buffer, &cpu_buffer).await;
+
+    // Ideally, we could assert here that the second compilation was faster than the first
+    // However, that doesn't actually work, because drivers have their own internal caches.
+    // This does work on my machine if I set `MESA_DISABLE_PIPELINE_CACHE=1`
+    // before running the test; but of course that is not a realistic scenario
+}
+
+async fn validate_pipeline(
+    ctx: &TestingContext,
+    pipeline: wgpu::ComputePipeline,
+    bind_group: &wgpu::BindGroup,
+    gpu_buffer: &wgpu::Buffer,
+    cpu_buffer: &wgpu::Buffer,
+) {
+    let mut encoder = ctx
+        .device
+        .create_command_encoder(&wgpu::CommandEncoderDescriptor {
+            label: Some("encoder"),
+        });
+
+    {
+        let mut cpass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
+            label: Some("compute_pass"),
+            timestamp_writes: None,
+        });
+        cpass.set_pipeline(&pipeline);
+        cpass.set_bind_group(0, bind_group, &[]);
+
+        cpass.dispatch_workgroups(1, 1, 1);
+    }
+
+    encoder.copy_buffer_to_buffer(gpu_buffer, 0, cpu_buffer, 0, ARRAY_SIZE * 4);
+    ctx.queue.submit([encoder.finish()]);
+    cpu_buffer.slice(..).map_async(wgpu::MapMode::Read, |_| ());
+    ctx.async_poll(wgpu::Maintain::wait())
+        .await
+        .panic_on_timeout();
+
+    let data = cpu_buffer.slice(..).get_mapped_range();
+
+    let arrays: &[u32] = bytemuck::cast_slice(&data);
+
+    assert_eq!(arrays.len(), ARRAY_SIZE as usize);
+    for (idx, value) in arrays.iter().copied().enumerate() {
+        assert_eq!(value as usize, idx);
+    }
+    drop(data);
+    cpu_buffer.unmap();
+}

tests/tests/push_constants.rs

@@ -104,6 +104,7 @@ async fn partial_update_test(ctx: TestingContext) {
             module: &sm,
             entry_point: "main",
             compilation_options: Default::default(),
+            cache: None,
         });
 
     let mut encoder = ctx

tests/tests/regression/issue_3349.rs

@@ -119,6 +119,7 @@ async fn multi_stage_data_binding_test(ctx: TestingContext) {
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
 
     let texture = ctx.device.create_texture(&wgpu::TextureDescriptor {

tests/tests/regression/issue_3457.rs

@@ -80,6 +80,7 @@ static PASS_RESET_VERTEX_BUFFER: GpuTestConfiguration =
                     })],
                 }),
                 multiview: None,
+                cache: None,
             });
 
         let single_pipeline = ctx
@@ -111,6 +112,7 @@ static PASS_RESET_VERTEX_BUFFER: GpuTestConfiguration =
                     })],
                 }),
                 multiview: None,
+                cache: None,
             });
 
         let view = ctx

tests/tests/root.rs

@@ -24,6 +24,7 @@ mod nv12_texture;
 mod occlusion_query;
 mod partially_bounded_arrays;
 mod pipeline;
+mod pipeline_cache;
 mod poll;
 mod push_constants;
 mod query_set;

tests/tests/scissor_tests/mod.rs

@@ -61,6 +61,7 @@ async fn scissor_test_impl(
                 })],
             }),
             multiview: None,
+            cache: None,
         });
 
     let readback_buffer = image::ReadbackBuffers::new(&ctx.device, &texture);

tests/tests/shader/mod.rs

@@ -310,6 +310,7 @@ async fn shader_input_output_test(
                 module: &sm,
                 entry_point: "cs_main",
                 compilation_options: Default::default(),
+                cache: None,
             });
 
         // -- Initializing data --

tests/tests/shader/zero_init_workgroup_mem.rs

@@ -88,6 +88,7 @@ static ZERO_INIT_WORKGROUP_MEMORY: GpuTestConfiguration = GpuTestConfiguration::
                 module: &sm,
                 entry_point: "read",
                 compilation_options: Default::default(),
+                cache: None,
             });
 
         let pipeline_write = ctx
@@ -98,6 +99,7 @@ static ZERO_INIT_WORKGROUP_MEMORY: GpuTestConfiguration = GpuTestConfiguration::
                 module: &sm,
                 entry_point: "write",
                 compilation_options: Default::default(),
+                cache: None,
             });
 
         // -- Initializing data --

tests/tests/shader_primitive_index/mod.rs

@@ -147,6 +147,7 @@ async fn pulling_common(
                 })],
             }),
             multiview: None,
+            cache: None,
         });
 
     let width = 2;

tests/tests/shader_view_format/mod.rs

@@ -109,6 +109,7 @@ async fn reinterpret(
             depth_stencil: None,
             multisample: wgpu::MultisampleState::default(),
             multiview: None,
+            cache: None,
         });
     let bind_group = ctx.device.create_bind_group(&wgpu::BindGroupDescriptor {
         layout: &pipeline.get_bind_group_layout(0),

tests/tests/subgroup_operations/mod.rs

@@ -75,6 +75,7 @@ static SUBGROUP_OPERATIONS: GpuTestConfiguration = GpuTestConfiguration::new()
             module: &cs_module,
             entry_point: "main",
             compilation_options: Default::default(),
+            cache: None,
         });
 
         let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {

tests/tests/vertex_indices/mod.rs

@@ -295,6 +295,7 @@ async fn vertex_index_common(ctx: TestingContext) {
             })],
         }),
         multiview: None,
+        cache: None,
     };
     let builtin_pipeline = ctx.device.create_render_pipeline(&pipeline_desc);
     pipeline_desc.vertex.entry_point = "vs_main_buffers";

wgpu-core/src/device/global.rs

@@ -13,8 +13,10 @@ use crate::{
     instance::{self, Adapter, Surface},
     lock::{rank, RwLock},
     pipeline, present,
-    resource::{self, BufferAccessResult},
-    resource::{BufferAccessError, BufferMapOperation, CreateBufferError, Resource},
+    resource::{
+        self, BufferAccessError, BufferAccessResult, BufferMapOperation, CreateBufferError,
+        Resource,
+    },
     validation::check_buffer_usage,
     Label, LabelHelpers as _,
 };
@@ -1823,6 +1825,66 @@ impl Global {
         }
     }
 
+    /// # Safety
+    /// The `data` argument of `desc` must have been returned by
+    /// [Self::pipeline_cache_get_data] for the same adapter
+    pub unsafe fn device_create_pipeline_cache<A: HalApi>(
+        &self,
+        device_id: DeviceId,
+        desc: &pipeline::PipelineCacheDescriptor<'_>,
+        id_in: Option<id::PipelineCacheId>,
+    ) -> (
+        id::PipelineCacheId,
+        Option<pipeline::CreatePipelineCacheError>,
+    ) {
+        profiling::scope!("Device::create_pipeline_cache");
+
+        let hub = A::hub(self);
+
+        let fid = hub.pipeline_caches.prepare(id_in);
+        let error: pipeline::CreatePipelineCacheError = 'error: {
+            let device = match hub.devices.get(device_id) {
+                Ok(device) => device,
+                // TODO: Handle error properly
+                Err(crate::storage::InvalidId) => break 'error DeviceError::Invalid.into(),
+            };
+            if !device.is_valid() {
+                break 'error DeviceError::Lost.into();
+            }
+            #[cfg(feature = "trace")]
+            if let Some(ref mut trace) = *device.trace.lock() {
+                trace.add(trace::Action::CreatePipelineCache {
+                    id: fid.id(),
+                    desc: desc.clone(),
+                });
+            }
+            let cache = unsafe { device.create_pipeline_cache(desc) };
+            match cache {
+                Ok(cache) => {
+                    let (id, _) = fid.assign(Arc::new(cache));
+                    api_log!("Device::create_pipeline_cache -> {id:?}");
+                    return (id, None);
+                }
+                Err(e) => break 'error e,
+            }
+        };
+
+        let id = fid.assign_error(desc.label.borrow_or_default());
+
+        (id, Some(error))
+    }
+
+    pub fn pipeline_cache_drop<A: HalApi>(&self, pipeline_cache_id: id::PipelineCacheId) {
+        profiling::scope!("PipelineCache::drop");
+        api_log!("PipelineCache::drop {pipeline_cache_id:?}");
+
+        let hub = A::hub(self);
+
+        if let Some(cache) = hub.pipeline_caches.unregister(pipeline_cache_id) {
+            drop(cache)
+        }
+    }
+
     pub fn surface_configure<A: HalApi>(
         &self,
         surface_id: SurfaceId,
@@ -2270,6 +2332,37 @@ impl Global {
             .force_replace_with_error(device_id, "Made invalid.");
     }
 
+    pub fn pipeline_cache_get_data<A: HalApi>(&self, id: id::PipelineCacheId) -> Option<Vec<u8>> {
+        use crate::pipeline_cache;
+        api_log!("PipelineCache::get_data");
+        let hub = A::hub(self);
+
+        if let Ok(cache) = hub.pipeline_caches.get(id) {
+            // TODO: Is this check needed?
+            if !cache.device.is_valid() {
+                return None;
+            }
+            if let Some(raw_cache) = cache.raw.as_ref() {
+                let mut vec = unsafe { cache.device.raw().pipeline_cache_get_data(raw_cache) }?;
+                let validation_key = cache.device.raw().pipeline_cache_validation_key()?;
+
+                let mut header_contents = [0; pipeline_cache::HEADER_LENGTH];
+                pipeline_cache::add_cache_header(
+                    &mut header_contents,
+                    &vec,
+                    &cache.device.adapter.raw.info,
+                    validation_key,
+                );
+
+                let deleted = vec.splice(..0, header_contents).collect::<Vec<_>>();
+                debug_assert!(deleted.is_empty());
+
+                return Some(vec);
+            }
+        }
+        None
+    }
+
     pub fn device_drop<A: HalApi>(&self, device_id: DeviceId) {
         profiling::scope!("Device::drop");
         api_log!("Device::drop {device_id:?}");

wgpu-core/src/device/resource.rs

@@ -2817,6 +2817,20 @@ impl<A: HalApi> Device<A> {
         let late_sized_buffer_groups =
             Device::make_late_sized_buffer_groups(&shader_binding_sizes, &pipeline_layout);
 
+        let cache = 'cache: {
+            let Some(cache) = desc.cache else {
+                break 'cache None;
+            };
+            let Ok(cache) = hub.pipeline_caches.get(cache) else {
+                break 'cache None;
+            };
+
+            if cache.device.as_info().id() != self.as_info().id() {
+                return Err(DeviceError::WrongDevice.into());
+            }
+            Some(cache)
+        };
+
         let pipeline_desc = hal::ComputePipelineDescriptor {
             label: desc.label.to_hal(self.instance_flags),
             layout: pipeline_layout.raw(),
@@ -2826,6 +2840,7 @@ impl<A: HalApi> Device<A> {
                 constants: desc.stage.constants.as_ref(),
                 zero_initialize_workgroup_memory: desc.stage.zero_initialize_workgroup_memory,
             },
+            cache: cache.as_ref().and_then(|it| it.raw.as_ref()),
         };
 
         let raw = unsafe {
@@ -3199,6 +3214,7 @@ impl<A: HalApi> Device<A> {
 
         let vertex_shader_module;
         let vertex_entry_point_name;
+
         let vertex_stage = {
             let stage_desc = &desc.vertex.stage;
             let stage = wgt::ShaderStages::VERTEX;
@@ -3393,6 +3409,20 @@ impl<A: HalApi> Device<A> {
         let late_sized_buffer_groups =
             Device::make_late_sized_buffer_groups(&shader_binding_sizes, &pipeline_layout);
 
+        let pipeline_cache = 'cache: {
+            let Some(cache) = desc.cache else {
+                break 'cache None;
+            };
+            let Ok(cache) = hub.pipeline_caches.get(cache) else {
+                break 'cache None;
+            };
+
+            if cache.device.as_info().id() != self.as_info().id() {
+                return Err(DeviceError::WrongDevice.into());
+            }
+            Some(cache)
+        };
+
         let pipeline_desc = hal::RenderPipelineDescriptor {
             label: desc.label.to_hal(self.instance_flags),
             layout: pipeline_layout.raw(),
@@ -3404,6 +3434,7 @@ impl<A: HalApi> Device<A> {
             fragment_stage,
             color_targets,
             multiview: desc.multiview,
+            cache: pipeline_cache.as_ref().and_then(|it| it.raw.as_ref()),
         };
         let raw = unsafe {
             self.raw
@@ -3484,6 +3515,53 @@ impl<A: HalApi> Device<A> {
         Ok(pipeline)
     }
 
+    /// # Safety
+    /// The `data` field on `desc` must have previously been returned from [`crate::global::Global::pipeline_cache_get_data`]
+    pub unsafe fn create_pipeline_cache(
+        self: &Arc<Self>,
+        desc: &pipeline::PipelineCacheDescriptor,
+    ) -> Result<pipeline::PipelineCache<A>, pipeline::CreatePipelineCacheError> {
+        use crate::pipeline_cache;
+        self.require_features(wgt::Features::PIPELINE_CACHE)?;
+        let data = if let Some((data, validation_key)) = desc
+            .data
+            .as_ref()
+            .zip(self.raw().pipeline_cache_validation_key())
+        {
+            let data = pipeline_cache::validate_pipeline_cache(
+                data,
+                &self.adapter.raw.info,
+                validation_key,
+            );
+            match data {
+                Ok(data) => Some(data),
+                Err(e) if e.was_avoidable() || !desc.fallback => return Err(e.into()),
+                // If the error was unavoidable and we are asked to fallback, do so
+                Err(_) => None,
+            }
+        } else {
+            None
+        };
+        let cache_desc = hal::PipelineCacheDescriptor {
+            data,
+            label: desc.label.to_hal(self.instance_flags),
+        };
+        let raw = match unsafe { self.raw().create_pipeline_cache(&cache_desc) } {
+            Ok(raw) => raw,
+            Err(e) => return Err(e.into()),
+        };
+        let cache = pipeline::PipelineCache {
+            device: self.clone(),
+            info: ResourceInfo::new(
+                desc.label.borrow_or_default(),
+                Some(self.tracker_indices.pipeline_caches.clone()),
+            ),
+            // This would be none in the error condition, which we don't implement yet
+            raw: Some(raw),
+        };
+        Ok(cache)
+    }
+
     pub(crate) fn get_texture_format_features(
         &self,
         adapter: &Adapter<A>,

wgpu-core/src/device/trace.rs

@@ -98,6 +98,11 @@ pub enum Action<'a> {
         implicit_context: Option<super::ImplicitPipelineContext>,
     },
     DestroyRenderPipeline(id::RenderPipelineId),
+    CreatePipelineCache {
+        id: id::PipelineCacheId,
+        desc: crate::pipeline::PipelineCacheDescriptor<'a>,
+    },
+    DestroyPipelineCache(id::PipelineCacheId),
     CreateRenderBundle {
         id: id::RenderBundleId,
         desc: crate::command::RenderBundleEncoderDescriptor<'a>,

wgpu-core/src/hub.rs

@@ -110,7 +110,7 @@ use crate::{
     device::{queue::Queue, Device},
     hal_api::HalApi,
     instance::{Adapter, Surface},
-    pipeline::{ComputePipeline, RenderPipeline, ShaderModule},
+    pipeline::{ComputePipeline, PipelineCache, RenderPipeline, ShaderModule},
     registry::{Registry, RegistryReport},
     resource::{Buffer, QuerySet, Sampler, StagingBuffer, Texture, TextureView},
     storage::{Element, Storage},
@@ -130,6 +130,7 @@ pub struct HubReport {
     pub render_bundles: RegistryReport,
     pub render_pipelines: RegistryReport,
     pub compute_pipelines: RegistryReport,
+    pub pipeline_caches: RegistryReport,
     pub query_sets: RegistryReport,
     pub buffers: RegistryReport,
     pub textures: RegistryReport,
@@ -180,6 +181,7 @@ pub struct Hub<A: HalApi> {
     pub(crate) render_bundles: Registry<RenderBundle<A>>,
     pub(crate) render_pipelines: Registry<RenderPipeline<A>>,
     pub(crate) compute_pipelines: Registry<ComputePipeline<A>>,
+    pub(crate) pipeline_caches: Registry<PipelineCache<A>>,
     pub(crate) query_sets: Registry<QuerySet<A>>,
     pub(crate) buffers: Registry<Buffer<A>>,
     pub(crate) staging_buffers: Registry<StagingBuffer<A>>,
@@ -202,6 +204,7 @@ impl<A: HalApi> Hub<A> {
             render_bundles: Registry::new(A::VARIANT),
             render_pipelines: Registry::new(A::VARIANT),
             compute_pipelines: Registry::new(A::VARIANT),
+            pipeline_caches: Registry::new(A::VARIANT),
             query_sets: Registry::new(A::VARIANT),
             buffers: Registry::new(A::VARIANT),
             staging_buffers: Registry::new(A::VARIANT),
@@ -235,6 +238,7 @@ impl<A: HalApi> Hub<A> {
         self.pipeline_layouts.write().map.clear();
         self.compute_pipelines.write().map.clear();
         self.render_pipelines.write().map.clear();
+        self.pipeline_caches.write().map.clear();
         self.query_sets.write().map.clear();
 
         for element in surface_guard.map.iter() {
@@ -280,6 +284,7 @@ impl<A: HalApi> Hub<A> {
             render_bundles: self.render_bundles.generate_report(),
             render_pipelines: self.render_pipelines.generate_report(),
             compute_pipelines: self.compute_pipelines.generate_report(),
+            pipeline_caches: self.pipeline_caches.generate_report(),
             query_sets: self.query_sets.generate_report(),
             buffers: self.buffers.generate_report(),
             textures: self.textures.generate_report(),

wgpu-core/src/id.rs

@@ -313,6 +313,7 @@ ids! {
     pub type ShaderModuleId ShaderModule;
     pub type RenderPipelineId RenderPipeline;
     pub type ComputePipelineId ComputePipeline;
+    pub type PipelineCacheId PipelineCache;
     pub type CommandEncoderId CommandEncoder;
     pub type CommandBufferId CommandBuffer;
     pub type RenderPassEncoderId RenderPassEncoder;

wgpu-core/src/lib.rs

@@ -65,6 +65,7 @@ mod init_tracker;
 pub mod instance;
 mod lock;
 pub mod pipeline;
+mod pipeline_cache;
 mod pool;
 pub mod present;
 pub mod registry;

wgpu-core/src/pipeline.rs

@@ -1,11 +1,12 @@
 #[cfg(feature = "trace")]
 use crate::device::trace;
+pub use crate::pipeline_cache::PipelineCacheValidationError;
 use crate::{
     binding_model::{CreateBindGroupLayoutError, CreatePipelineLayoutError, PipelineLayout},
     command::ColorAttachmentError,
     device::{Device, DeviceError, MissingDownlevelFlags, MissingFeatures, RenderPassContext},
     hal_api::HalApi,
-    id::{PipelineLayoutId, ShaderModuleId},
+    id::{PipelineCacheId, PipelineLayoutId, ShaderModuleId},
     resource::{Resource, ResourceInfo, ResourceType},
     resource_log, validation, Label,
 };
@@ -192,6 +193,8 @@ pub struct ComputePipelineDescriptor<'a> {
     pub layout: Option<PipelineLayoutId>,
     /// The compiled compute stage and its entry point.
     pub stage: ProgrammableStageDescriptor<'a>,
+    /// The pipeline cache to use when creating this pipeline.
+    pub cache: Option<PipelineCacheId>,
 }
 
 #[derive(Clone, Debug, Error)]
@@ -259,6 +262,68 @@ impl<A: HalApi> ComputePipeline<A> {
     }
 }
 
+#[derive(Clone, Debug, Error)]
+#[non_exhaustive]
+pub enum CreatePipelineCacheError {
+    #[error(transparent)]
+    Device(#[from] DeviceError),
+    #[error("Pipeline cache validation failed")]
+    Validation(#[from] PipelineCacheValidationError),
+    #[error(transparent)]
+    MissingFeatures(#[from] MissingFeatures),
+    #[error("Internal error: {0}")]
+    Internal(String),
+}
+
+impl From<hal::PipelineCacheError> for CreatePipelineCacheError {
+    fn from(value: hal::PipelineCacheError) -> Self {
+        match value {
+            hal::PipelineCacheError::Device(device) => {
+                CreatePipelineCacheError::Device(device.into())
+            }
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct PipelineCache<A: HalApi> {
+    pub(crate) raw: Option<A::PipelineCache>,
+    pub(crate) device: Arc<Device<A>>,
+    pub(crate) info: ResourceInfo<PipelineCache<A>>,
+}
+
+impl<A: HalApi> Drop for PipelineCache<A> {
+    fn drop(&mut self) {
+        if let Some(raw) = self.raw.take() {
+            resource_log!("Destroy raw PipelineCache {:?}", self.info.label());
+
+            #[cfg(feature = "trace")]
+            if let Some(t) = self.device.trace.lock().as_mut() {
+                t.add(trace::Action::DestroyPipelineCache(self.info.id()));
+            }
+
+            unsafe {
+                use hal::Device;
+                self.device.raw().destroy_pipeline_cache(raw);
+            }
+        }
+    }
+}
+
+impl<A: HalApi> Resource for PipelineCache<A> {
+    const TYPE: ResourceType = "PipelineCache";
+
+    type Marker = crate::id::markers::PipelineCache;
+
+    fn as_info(&self) -> &ResourceInfo<Self> {
+        &self.info
+    }
+
+    fn as_info_mut(&mut self) -> &mut ResourceInfo<Self> {
+        &mut self.info
+    }
+}
+
 /// Describes how the vertex buffer is interpreted.
 #[derive(Clone, Debug)]
 #[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
@@ -315,6 +380,16 @@ pub struct RenderPipelineDescriptor<'a> {
     /// If the pipeline will be used with a multiview render pass, this indicates how many array
     /// layers the attachments will have.
     pub multiview: Option<NonZeroU32>,
+    /// The pipeline cache to use when creating this pipeline.
+    pub cache: Option<PipelineCacheId>,
+}
+
+#[derive(Clone, Debug)]
+#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
+pub struct PipelineCacheDescriptor<'a> {
+    pub label: Label<'a>,
+    pub data: Option<Cow<'a, [u8]>>,
+    pub fallback: bool,
 }
 
 #[derive(Clone, Debug, Error)]

wgpu-core/src/pipeline_cache.rs

@@ -0,0 +1,530 @@
+use thiserror::Error;
+use wgt::AdapterInfo;
+
+pub const HEADER_LENGTH: usize = std::mem::size_of::<PipelineCacheHeader>();
+
+#[derive(Debug, PartialEq, Eq, Clone, Error)]
+#[non_exhaustive]
+pub enum PipelineCacheValidationError {
+    #[error("The pipeline cache data was truncated")]
+    Truncated,
+    #[error("The pipeline cache data was longer than recorded")]
+    // TODO: Is it plausible that this would happen
+    Extended,
+    #[error("The pipeline cache data was corrupted (e.g. the hash didn't match)")]
+    Corrupted,
+    #[error("The pipeline cacha data was out of date and so cannot be safely used")]
+    Outdated,
+    #[error("The cache data was created for a different device")]
+    WrongDevice,
+    #[error("Pipeline cacha data was created for a future version of wgpu")]
+    Unsupported,
+}
+
+impl PipelineCacheValidationError {
+    /// Could the error have been avoided?
+    /// That is, is there a mistake in user code interacting with the cache
+    pub fn was_avoidable(&self) -> bool {
+        match self {
+            PipelineCacheValidationError::WrongDevice => true,
+            PipelineCacheValidationError::Truncated
+            | PipelineCacheValidationError::Unsupported
+            | PipelineCacheValidationError::Extended
+            // It's unusual, but not implausible, to be downgrading wgpu
+            | PipelineCacheValidationError::Outdated
+            | PipelineCacheValidationError::Corrupted => false,
+        }
+    }
+}
+
+/// Validate the data in a pipeline cache
+pub fn validate_pipeline_cache<'d>(
+    cache_data: &'d [u8],
+    adapter: &AdapterInfo,
+    validation_key: [u8; 16],
+) -> Result<&'d [u8], PipelineCacheValidationError> {
+    let adapter_key = adapter_key(adapter)?;
+    let Some((header, remaining_data)) = PipelineCacheHeader::read(cache_data) else {
+        return Err(PipelineCacheValidationError::Truncated);
+    };
+    if header.magic != MAGIC {
+        return Err(PipelineCacheValidationError::Corrupted);
+    }
+    if header.header_version != HEADER_VERSION {
+        return Err(PipelineCacheValidationError::Outdated);
+    }
+    if header.cache_abi != ABI {
+        return Err(PipelineCacheValidationError::Outdated);
+    }
+    if header.backend != adapter.backend as u8 {
+        return Err(PipelineCacheValidationError::WrongDevice);
+    }
+    if header.adapter_key != adapter_key {
+        return Err(PipelineCacheValidationError::WrongDevice);
+    }
+    if header.validation_key != validation_key {
+        // If the validation key is wrong, that means that this device has changed
+        // in a way where the cache won't be compatible since the cache was made,
+        // so it is outdated
+        return Err(PipelineCacheValidationError::Outdated);
+    }
+    let data_size: usize = header
+        .data_size
+        .try_into()
+        // If the data was previously more than 4GiB, and we're still on a 32 bit system (ABI check, above)
+        // Then the data must be corrupted
+        .map_err(|_| PipelineCacheValidationError::Corrupted)?;
+    if remaining_data.len() < data_size {
+        return Err(PipelineCacheValidationError::Truncated);
+    }
+    if remaining_data.len() > data_size {
+        return Err(PipelineCacheValidationError::Extended);
+    }
+    if header.hash_space != HASH_SPACE_VALUE {
+        return Err(PipelineCacheValidationError::Corrupted);
+    }
+    Ok(remaining_data)
+}
+
+pub fn add_cache_header(
+    in_region: &mut [u8],
+    data: &[u8],
+    adapter: &AdapterInfo,
+    validation_key: [u8; 16],
+) {
+    assert_eq!(in_region.len(), HEADER_LENGTH);
+    let header = PipelineCacheHeader {
+        adapter_key: adapter_key(adapter)
+            .expect("Called add_cache_header for an adapter which doesn't support cache data. This is a wgpu internal bug"),
+        backend: adapter.backend as u8,
+        cache_abi: ABI,
+        magic: MAGIC,
+        header_version: HEADER_VERSION,
+        validation_key,
+        hash_space: HASH_SPACE_VALUE,
+        data_size: data
+            .len()
+            .try_into()
+            .expect("Cache larger than u64::MAX bytes"),
+    };
+    header.write(in_region);
+}
+
+const MAGIC: [u8; 8] = *b"WGPUPLCH";
+const HEADER_VERSION: u32 = 1;
+const ABI: u32 = std::mem::size_of::<*const ()>() as u32;
+
+/// The value used to fill [`PipelineCacheHeader::hash_space`]
+///
+/// If we receive reports of pipeline cache data corruption which is not otherwise caught
+/// on a real device, it would be worth modifying this
+///
+/// Note that wgpu does not protect against malicious writes to e.g. a file used
+/// to store a pipeline cache.
+/// That is the resonsibility of the end application, such as by using a
+/// private space.
+const HASH_SPACE_VALUE: u64 = 0xFEDCBA9_876543210;
+
+#[repr(C)]
+#[derive(PartialEq, Eq)]
+struct PipelineCacheHeader {
+    /// The magic header to ensure that we have the right file format
+    /// Has a value of MAGIC, as above
+    magic: [u8; 8],
+    // /// The total size of this header, in bytes
+    // header_size: u32,
+    /// The version of this wgpu header
+    /// Should be equal to HEADER_VERSION above
+    ///
+    /// This must always be the second item, after the value above
+    header_version: u32,
+    /// The number of bytes in the pointers of this ABI, because some drivers
+    /// have previously not distinguished between their 32 bit and 64 bit drivers
+    /// leading to Vulkan data corruption
+    cache_abi: u32,
+    /// The id for the backend in use, from [wgt::Backend]
+    backend: u8,
+    /// The key which identifiers the device/adapter.
+    /// This is used to validate that this pipeline cache (probably) was produced for
+    /// the expected device.
+    /// On Vulkan: it is a combination of vendor ID and device ID
+    adapter_key: [u8; 15],
+    /// A key used to validate that this device is still compatible with the cache
+    ///
+    /// This should e.g. contain driver version and/or intermediate compiler versions
+    validation_key: [u8; 16],
+    /// The length of the data which is sent to/recieved from the backend
+    data_size: u64,
+    /// Space reserved for a hash of the data in future
+    ///
+    /// We assume that your cache storage system will be relatively robust, and so
+    /// do not validate this hash
+    ///
+    /// Therefore, this will always have a value of [`HASH_SPACE_VALUE`]
+    hash_space: u64,
+}
+
+impl PipelineCacheHeader {
+    fn read(data: &[u8]) -> Option<(PipelineCacheHeader, &[u8])> {
+        let mut reader = Reader {
+            data,
+            total_read: 0,
+        };
+        let magic = reader.read_array()?;
+        let header_version = reader.read_u32()?;
+        let cache_abi = reader.read_u32()?;
+        let backend = reader.read_byte()?;
+        let adapter_key = reader.read_array()?;
+        let validation_key = reader.read_array()?;
+        let data_size = reader.read_u64()?;
+        let data_hash = reader.read_u64()?;
+
+        assert_eq!(
+            reader.total_read,
+            std::mem::size_of::<PipelineCacheHeader>()
+        );
+
+        Some((
+            PipelineCacheHeader {
+                magic,
+                header_version,
+                cache_abi,
+                backend,
+                adapter_key,
+                validation_key,
+                data_size,
+                hash_space: data_hash,
+            },
+            reader.data,
+        ))
+    }
+
+    fn write(&self, into: &mut [u8]) -> Option<()> {
+        let mut writer = Writer { data: into };
+        writer.write_array(&self.magic)?;
+        writer.write_u32(self.header_version)?;
+        writer.write_u32(self.cache_abi)?;
+        writer.write_byte(self.backend)?;
+        writer.write_array(&self.adapter_key)?;
+        writer.write_array(&self.validation_key)?;
+        writer.write_u64(self.data_size)?;
+        writer.write_u64(self.hash_space)?;
+
+        assert_eq!(writer.data.len(), 0);
+        Some(())
+    }
+}
+
+fn adapter_key(adapter: &AdapterInfo) -> Result<[u8; 15], PipelineCacheValidationError> {
+    match adapter.backend {
+        wgt::Backend::Vulkan => {
+            // If these change size, the header format needs to change
+            // We set the type explicitly so this won't compile in that case
+            let v: [u8; 4] = adapter.vendor.to_be_bytes();
+            let d: [u8; 4] = adapter.device.to_be_bytes();
+            let adapter = [
+                255, 255, 255, v[0], v[1], v[2], v[3], d[0], d[1], d[2], d[3], 255, 255, 255, 255,
+            ];
+            Ok(adapter)
+        }
+        _ => Err(PipelineCacheValidationError::Unsupported),
+    }
+}
+
+struct Reader<'a> {
+    data: &'a [u8],
+    total_read: usize,
+}
+
+impl<'a> Reader<'a> {
+    fn read_byte(&mut self) -> Option<u8> {
+        let res = *self.data.first()?;
+        self.total_read += 1;
+        self.data = &self.data[1..];
+        Some(res)
+    }
+    fn read_array<const N: usize>(&mut self) -> Option<[u8; N]> {
+        // Only greater than because we're indexing fenceposts, not items
+        if N > self.data.len() {
+            return None;
+        }
+        let (start, data) = self.data.split_at(N);
+        self.total_read += N;
+        self.data = data;
+        Some(start.try_into().expect("off-by-one-error in array size"))
+    }
+
+    // fn read_u16(&mut self) -> Option<u16> {
+    //     self.read_array().map(u16::from_be_bytes)
+    // }
+    fn read_u32(&mut self) -> Option<u32> {
+        self.read_array().map(u32::from_be_bytes)
+    }
+    fn read_u64(&mut self) -> Option<u64> {
+        self.read_array().map(u64::from_be_bytes)
+    }
+}
+
+struct Writer<'a> {
+    data: &'a mut [u8],
+}
+
+impl<'a> Writer<'a> {
+    fn write_byte(&mut self, byte: u8) -> Option<()> {
+        self.write_array(&[byte])
+    }
+    fn write_array<const N: usize>(&mut self, array: &[u8; N]) -> Option<()> {
+        // Only greater than because we're indexing fenceposts, not items
+        if N > self.data.len() {
+            return None;
+        }
+        let data = std::mem::take(&mut self.data);
+        let (start, data) = data.split_at_mut(N);
+        self.data = data;
+        start.copy_from_slice(array);
+        Some(())
+    }
+
+    // fn write_u16(&mut self, value: u16) -> Option<()> {
+    //     self.write_array(&value.to_be_bytes())
+    // }
+    fn write_u32(&mut self, value: u32) -> Option<()> {
+        self.write_array(&value.to_be_bytes())
+    }
+    fn write_u64(&mut self, value: u64) -> Option<()> {
+        self.write_array(&value.to_be_bytes())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use wgt::AdapterInfo;
+
+    use crate::pipeline_cache::{PipelineCacheValidationError as E, HEADER_LENGTH};
+
+    use super::ABI;
+
+    // Assert the correct size
+    const _: [(); HEADER_LENGTH] = [(); 64];
+
+    const ADAPTER: AdapterInfo = AdapterInfo {
+        name: String::new(),
+        vendor: 0x0002_FEED,
+        device: 0xFEFE_FEFE,
+        device_type: wgt::DeviceType::Other,
+        driver: String::new(),
+        driver_info: String::new(),
+        backend: wgt::Backend::Vulkan,
+    };
+
+    // IMPORTANT: If these tests fail, then you MUST increment HEADER_VERSION
+    const VALIDATION_KEY: [u8; 16] = u128::to_be_bytes(0xFFFFFFFF_FFFFFFFF_88888888_88888888);
+    #[test]
+    fn written_header() {
+        let mut result = [0; HEADER_LENGTH];
+        super::add_cache_header(&mut result, &[], &ADAPTER, VALIDATION_KEY);
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let expected = cache.into_iter().flatten().collect::<Vec<u8>>();
+
+        assert_eq!(result.as_slice(), expected.as_slice());
+    }
+
+    #[test]
+    fn valid_data() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let expected: &[u8] = &[];
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Ok(expected));
+    }
+    #[test]
+    fn invalid_magic() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"NOT_WGPU",                                 // (Wrong) MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Corrupted));
+    }
+
+    #[test]
+    fn wrong_version() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 2, 0, 0, 0, ABI as u8],             // (wrong) Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Outdated));
+    }
+    #[test]
+    fn wrong_abi() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH", // MAGIC
+            // a 14 bit ABI is improbable
+            [0, 0, 0, 1, 0, 0, 0, 14],            // Version and (wrong) ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED], // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(), // Validation key
+            0x88888888_88888888u64.to_be_bytes(), // Validation key
+            0x0u64.to_be_bytes(),                 // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(), // Header
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Outdated));
+    }
+
+    #[test]
+    fn wrong_backend() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [2, 255, 255, 255, 0, 2, 0xFE, 0xED],         // (wrong) Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::WrongDevice));
+    }
+    #[test]
+    fn wrong_adapter() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0x00],         // Backend and (wrong) Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::WrongDevice));
+    }
+    #[test]
+    fn wrong_validation() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_00000000u64.to_be_bytes(),         // (wrong) Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Outdated));
+    }
+    #[test]
+    fn too_little_data() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x064u64.to_be_bytes(),                       // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Truncated));
+    }
+    #[test]
+    fn not_no_data() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            100u64.to_be_bytes(),                         // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache
+            .into_iter()
+            .flatten()
+            .chain(std::iter::repeat(0u8).take(100))
+            .collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        let expected: &[u8] = &[0; 100];
+        assert_eq!(validation_result, Ok(expected));
+    }
+    #[test]
+    fn too_much_data() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x064u64.to_be_bytes(),                       // Data size
+            0xFEDCBA9_876543210u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache
+            .into_iter()
+            .flatten()
+            .chain(std::iter::repeat(0u8).take(200))
+            .collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Extended));
+    }
+    #[test]
+    fn wrong_hash() {
+        let cache: [[u8; 8]; HEADER_LENGTH / 8] = [
+            *b"WGPUPLCH",                                 // MAGIC
+            [0, 0, 0, 1, 0, 0, 0, ABI as u8],             // Version and ABI
+            [1, 255, 255, 255, 0, 2, 0xFE, 0xED],         // Backend and Adapter key
+            [0xFE, 0xFE, 0xFE, 0xFE, 255, 255, 255, 255], // Backend and Adapter key
+            0xFFFFFFFF_FFFFFFFFu64.to_be_bytes(),         // Validation key
+            0x88888888_88888888u64.to_be_bytes(),         // Validation key
+            0x0u64.to_be_bytes(),                         // Data size
+            0x00000000_00000000u64.to_be_bytes(),         // Hash
+        ];
+        let cache = cache.into_iter().flatten().collect::<Vec<u8>>();
+        let validation_result = super::validate_pipeline_cache(&cache, &ADAPTER, VALIDATION_KEY);
+        assert_eq!(validation_result, Err(E::Corrupted));
+    }
+}

wgpu-core/src/track/mod.rs

@@ -228,6 +228,7 @@ pub(crate) struct TrackerIndexAllocators {
     pub pipeline_layouts: Arc<SharedTrackerIndexAllocator>,
     pub bundles: Arc<SharedTrackerIndexAllocator>,
     pub query_sets: Arc<SharedTrackerIndexAllocator>,
+    pub pipeline_caches: Arc<SharedTrackerIndexAllocator>,
 }
 
 impl TrackerIndexAllocators {
@@ -245,6 +246,7 @@ impl TrackerIndexAllocators {
             pipeline_layouts: Arc::new(SharedTrackerIndexAllocator::new()),
             bundles: Arc::new(SharedTrackerIndexAllocator::new()),
             query_sets: Arc::new(SharedTrackerIndexAllocator::new()),
+            pipeline_caches: Arc::new(SharedTrackerIndexAllocator::new()),
         }
     }
 }

wgpu-hal/examples/halmark/main.rs

@@ -274,6 +274,7 @@ impl<A: hal::Api> Example<A> {
                 write_mask: wgt::ColorWrites::default(),
             })],
             multiview: None,
+            cache: None,
         };
         let pipeline = unsafe { device.create_render_pipeline(&pipeline_desc).unwrap() };
 

wgpu-hal/examples/ray-traced-triangle/main.rs

@@ -374,6 +374,7 @@ impl<A: hal::Api> Example<A> {
                     constants: &Default::default(),
                     zero_initialize_workgroup_memory: true,
                 },
+                cache: None,
             })
         }
         .unwrap();

wgpu-hal/src/dx12/device.rs

@@ -1513,6 +1513,14 @@ impl crate::Device for super::Device {
     }
     unsafe fn destroy_compute_pipeline(&self, _pipeline: super::ComputePipeline) {}
 
+    unsafe fn create_pipeline_cache(
+        &self,
+        _desc: &crate::PipelineCacheDescriptor<'_>,
+    ) -> Result<(), crate::PipelineCacheError> {
+        Ok(())
+    }
+    unsafe fn destroy_pipeline_cache(&self, (): ()) {}
+
     unsafe fn create_query_set(
         &self,
         desc: &wgt::QuerySetDescriptor<crate::Label>,

wgpu-hal/src/dx12/mod.rs

@@ -82,6 +82,7 @@ impl crate::Api for Api {
     type ShaderModule = ShaderModule;
     type RenderPipeline = RenderPipeline;
     type ComputePipeline = ComputePipeline;
+    type PipelineCache = ();
 
     type AccelerationStructure = AccelerationStructure;
 }

wgpu-hal/src/empty.rs

@@ -30,6 +30,7 @@ impl crate::Api for Api {
     type QuerySet = Resource;
     type Fence = Resource;
     type AccelerationStructure = Resource;
+    type PipelineCache = Resource;
 
     type BindGroupLayout = Resource;
     type BindGroup = Resource;
@@ -220,6 +221,13 @@ impl crate::Device for Context {
         Ok(Resource)
     }
     unsafe fn destroy_compute_pipeline(&self, pipeline: Resource) {}
+    unsafe fn create_pipeline_cache(
+        &self,
+        desc: &crate::PipelineCacheDescriptor<'_>,
+    ) -> Result<Resource, crate::PipelineCacheError> {
+        Ok(Resource)
+    }
+    unsafe fn destroy_pipeline_cache(&self, cache: Resource) {}
 
     unsafe fn create_query_set(
         &self,

wgpu-hal/src/gles/device.rs

@@ -1406,6 +1406,16 @@ impl crate::Device for super::Device {
         }
     }
 
+    unsafe fn create_pipeline_cache(
+        &self,
+        _: &crate::PipelineCacheDescriptor<'_>,
+    ) -> Result<(), crate::PipelineCacheError> {
+        // Even though the cache doesn't do anything, we still return something here
+        // as the least bad option
+        Ok(())
+    }
+    unsafe fn destroy_pipeline_cache(&self, (): ()) {}
+
     #[cfg_attr(target_arch = "wasm32", allow(unused))]
     unsafe fn create_query_set(
         &self,

wgpu-hal/src/gles/mod.rs

@@ -154,6 +154,7 @@ impl crate::Api for Api {
     type QuerySet = QuerySet;
     type Fence = Fence;
     type AccelerationStructure = ();
+    type PipelineCache = ();
 
     type BindGroupLayout = BindGroupLayout;
     type BindGroup = BindGroup;

wgpu-hal/src/lib.rs

@@ -332,6 +332,12 @@ pub enum PipelineError {
     Device(#[from] DeviceError),
 }
 
+#[derive(Clone, Debug, Eq, PartialEq, Error)]
+pub enum PipelineCacheError {
+    #[error(transparent)]
+    Device(#[from] DeviceError),
+}
+
 #[derive(Clone, Debug, Eq, PartialEq, Error)]
 pub enum SurfaceError {
     #[error("Surface is lost")]
@@ -432,6 +438,7 @@ pub trait Api: Clone + fmt::Debug + Sized {
     type ShaderModule: fmt::Debug + WasmNotSendSync;
     type RenderPipeline: fmt::Debug + WasmNotSendSync;
     type ComputePipeline: fmt::Debug + WasmNotSendSync;
+    type PipelineCache: fmt::Debug + WasmNotSendSync;
 
     type AccelerationStructure: fmt::Debug + WasmNotSendSync + 'static;
 }
@@ -611,6 +618,14 @@ pub trait Device: WasmNotSendSync {
         desc: &ComputePipelineDescriptor<Self::A>,
     ) -> Result<<Self::A as Api>::ComputePipeline, PipelineError>;
     unsafe fn destroy_compute_pipeline(&self, pipeline: <Self::A as Api>::ComputePipeline);
+    unsafe fn create_pipeline_cache(
+        &self,
+        desc: &PipelineCacheDescriptor<'_>,
+    ) -> Result<<Self::A as Api>::PipelineCache, PipelineCacheError>;
+    fn pipeline_cache_validation_key(&self) -> Option<[u8; 16]> {
+        None
+    }
+    unsafe fn destroy_pipeline_cache(&self, cache: <Self::A as Api>::PipelineCache);
 
     unsafe fn create_query_set(
         &self,
@@ -652,6 +667,14 @@ pub trait Device: WasmNotSendSync {
     unsafe fn start_capture(&self) -> bool;
     unsafe fn stop_capture(&self);
 
+    #[allow(unused_variables)]
+    unsafe fn pipeline_cache_get_data(
+        &self,
+        cache: &<Self::A as Api>::PipelineCache,
+    ) -> Option<Vec<u8>> {
+        None
+    }
+
     unsafe fn create_acceleration_structure(
         &self,
         desc: &AccelerationStructureDescriptor,
@@ -1636,6 +1659,13 @@ pub struct ComputePipelineDescriptor<'a, A: Api> {
     pub layout: &'a A::PipelineLayout,
     /// The compiled compute stage and its entry point.
     pub stage: ProgrammableStage<'a, A>,
+    /// The cache which will be used and filled when compiling this pipeline
+    pub cache: Option<&'a A::PipelineCache>,
+}
+
+pub struct PipelineCacheDescriptor<'a> {
+    pub label: Label<'a>,
+    pub data: Option<&'a [u8]>,
 }
 
 /// Describes how the vertex buffer is interpreted.
@@ -1672,6 +1702,8 @@ pub struct RenderPipelineDescriptor<'a, A: Api> {
     /// If the pipeline will be used with a multiview render pass, this indicates how many array
     /// layers the attachments will have.
     pub multiview: Option<NonZeroU32>,
+    /// The cache which will be used and filled when compiling this pipeline
+    pub cache: Option<&'a A::PipelineCache>,
 }
 
 #[derive(Debug, Clone)]

wgpu-hal/src/metal/device.rs

@@ -1099,6 +1099,14 @@ impl crate::Device for super::Device {
     }
     unsafe fn destroy_compute_pipeline(&self, _pipeline: super::ComputePipeline) {}
 
+    unsafe fn create_pipeline_cache(
+        &self,
+        _desc: &crate::PipelineCacheDescriptor<'_>,
+    ) -> Result<(), crate::PipelineCacheError> {
+        Ok(())
+    }
+    unsafe fn destroy_pipeline_cache(&self, (): ()) {}
+
     unsafe fn create_query_set(
         &self,
         desc: &wgt::QuerySetDescriptor<crate::Label>,

wgpu-hal/src/metal/mod.rs

@@ -66,6 +66,7 @@ impl crate::Api for Api {
     type ShaderModule = ShaderModule;
     type RenderPipeline = RenderPipeline;
     type ComputePipeline = ComputePipeline;
+    type PipelineCache = ();
 
     type AccelerationStructure = AccelerationStructure;
 }

wgpu-hal/src/vulkan/adapter.rs

@@ -462,7 +462,8 @@ impl PhysicalDeviceFeatures {
             | F::TIMESTAMP_QUERY_INSIDE_ENCODERS
             | F::TIMESTAMP_QUERY_INSIDE_PASSES
             | F::TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES
-            | F::CLEAR_TEXTURE;
+            | F::CLEAR_TEXTURE
+            | F::PIPELINE_CACHE;
 
         let mut dl_flags = Df::COMPUTE_SHADERS
             | Df::BASE_VERTEX
@@ -1745,6 +1746,19 @@ impl super::Adapter {
             unsafe { raw_device.get_device_queue(family_index, queue_index) }
         };
 
+        let driver_version = self
+            .phd_capabilities
+            .properties
+            .driver_version
+            .to_be_bytes();
+        #[rustfmt::skip]
+        let pipeline_cache_validation_key = [
+            driver_version[0], driver_version[1], driver_version[2], driver_version[3],
+            0, 0, 0, 0,
+            0, 0, 0, 0,
+            0, 0, 0, 0,
+        ];
+
         let shared = Arc::new(super::DeviceShared {
             raw: raw_device,
             family_index,
@@ -1760,6 +1774,7 @@ impl super::Adapter {
                 timeline_semaphore: timeline_semaphore_fn,
                 ray_tracing: ray_tracing_fns,
             },
+            pipeline_cache_validation_key,
             vendor_id: self.phd_capabilities.properties.vendor_id,
             timestamp_period: self.phd_capabilities.properties.limits.timestamp_period,
             private_caps: self.private_caps.clone(),

wgpu-hal/src/vulkan/device.rs

@@ -1,4 +1,4 @@
-use super::conv;
+use super::{conv, PipelineCache};
 
 use arrayvec::ArrayVec;
 use ash::{khr, vk};
@@ -1867,12 +1867,17 @@ impl crate::Device for super::Device {
                 .render_pass(raw_pass)
         }];
 
+        let pipeline_cache = desc
+            .cache
+            .map(|it| it.raw)
+            .unwrap_or(vk::PipelineCache::null());
+
         let mut raw_vec = {
             profiling::scope!("vkCreateGraphicsPipelines");
             unsafe {
                 self.shared
                     .raw
-                    .create_graphics_pipelines(vk::PipelineCache::null(), &vk_infos, None)
+                    .create_graphics_pipelines(pipeline_cache, &vk_infos, None)
                     .map_err(|(_, e)| crate::DeviceError::from(e))
             }?
         };
@@ -1915,12 +1920,17 @@ impl crate::Device for super::Device {
                 .stage(compiled.create_info)
         }];
 
+        let pipeline_cache = desc
+            .cache
+            .map(|it| it.raw)
+            .unwrap_or(vk::PipelineCache::null());
+
         let mut raw_vec = {
             profiling::scope!("vkCreateComputePipelines");
             unsafe {
                 self.shared
                     .raw
-                    .create_compute_pipelines(vk::PipelineCache::null(), &vk_infos, None)
+                    .create_compute_pipelines(pipeline_cache, &vk_infos, None)
                     .map_err(|(_, e)| crate::DeviceError::from(e))
             }?
         };
@@ -1940,6 +1950,26 @@ impl crate::Device for super::Device {
         unsafe { self.shared.raw.destroy_pipeline(pipeline.raw, None) };
     }
 
+    unsafe fn create_pipeline_cache(
+        &self,
+        desc: &crate::PipelineCacheDescriptor<'_>,
+    ) -> Result<PipelineCache, crate::PipelineCacheError> {
+        let mut info = vk::PipelineCacheCreateInfo::default();
+        if let Some(data) = desc.data {
+            info = info.initial_data(data)
+        }
+        profiling::scope!("vkCreatePipelineCache");
+        let raw = unsafe { self.shared.raw.create_pipeline_cache(&info, None) }
+            .map_err(crate::DeviceError::from)?;
+
+        Ok(PipelineCache { raw })
+    }
+    fn pipeline_cache_validation_key(&self) -> Option<[u8; 16]> {
+        Some(self.shared.pipeline_cache_validation_key)
+    }
+    unsafe fn destroy_pipeline_cache(&self, cache: PipelineCache) {
+        unsafe { self.shared.raw.destroy_pipeline_cache(cache.raw, None) }
+    }
     unsafe fn create_query_set(
         &self,
         desc: &wgt::QuerySetDescriptor<crate::Label>,
@@ -2105,6 +2135,11 @@ impl crate::Device for super::Device {
         }
     }
 
+    unsafe fn pipeline_cache_get_data(&self, cache: &PipelineCache) -> Option<Vec<u8>> {
+        let data = unsafe { self.raw_device().get_pipeline_cache_data(cache.raw) };
+        data.ok()
+    }
+
     unsafe fn get_acceleration_structure_build_sizes<'a>(
         &self,
         desc: &crate::GetAccelerationStructureBuildSizesDescriptor<'a, super::Api>,

wgpu-hal/src/vulkan/mod.rs

@@ -70,6 +70,7 @@ impl crate::Api for Api {
     type QuerySet = QuerySet;
     type Fence = Fence;
     type AccelerationStructure = AccelerationStructure;
+    type PipelineCache = PipelineCache;
 
     type BindGroupLayout = BindGroupLayout;
     type BindGroup = BindGroup;
@@ -338,6 +339,7 @@ struct DeviceShared {
     enabled_extensions: Vec<&'static CStr>,
     extension_fns: DeviceExtensionFunctions,
     vendor_id: u32,
+    pipeline_cache_validation_key: [u8; 16],
     timestamp_period: f32,
     private_caps: PrivateCapabilities,
     workarounds: Workarounds,
@@ -549,6 +551,11 @@ pub struct ComputePipeline {
     raw: vk::Pipeline,
 }
 
+#[derive(Debug)]
+pub struct PipelineCache {
+    raw: vk::PipelineCache,
+}
+
 #[derive(Debug)]
 pub struct QuerySet {
     raw: vk::QueryPool,

wgpu-types/src/lib.rs

@@ -914,6 +914,15 @@ bitflags::bitflags! {
         ///
         /// This is a native only feature.
         const SUBGROUP_BARRIER = 1 << 58;
+        /// Allows the use of pipeline cache objects
+        ///
+        /// Supported platforms:
+        /// - Vulkan
+        ///
+        /// Unimplemented Platforms:
+        /// - DX12
+        /// - Metal
+        const PIPELINE_CACHE = 1 << 59;
     }
 }
 

wgpu/src/backend/webgpu.rs

@@ -1159,6 +1159,8 @@ impl crate::context::Context for ContextWebGpu {
     type SurfaceOutputDetail = SurfaceOutputDetail;
     type SubmissionIndex = Unused;
     type SubmissionIndexData = ();
+    type PipelineCacheId = Unused;
+    type PipelineCacheData = ();
 
     type RequestAdapterFuture = MakeSendFuture<
         wasm_bindgen_futures::JsFuture,
@@ -1995,6 +1997,16 @@ impl crate::context::Context for ContextWebGpu {
         create_identified(device_data.0.create_compute_pipeline(&mapped_desc))
     }
 
+    unsafe fn device_create_pipeline_cache(
+        &self,
+        _: &Self::DeviceId,
+        _: &Self::DeviceData,
+        _: &crate::PipelineCacheDescriptor<'_>,
+    ) -> (Self::PipelineCacheId, Self::PipelineCacheData) {
+        (Unused, ())
+    }
+    fn pipeline_cache_drop(&self, _: &Self::PipelineCacheId, _: &Self::PipelineCacheData) {}
+
     fn device_create_buffer(
         &self,
         _device: &Self::DeviceId,
@@ -2981,6 +2993,14 @@ impl crate::context::Context for ContextWebGpu {
     fn device_start_capture(&self, _device: &Self::DeviceId, _device_data: &Self::DeviceData) {}
     fn device_stop_capture(&self, _device: &Self::DeviceId, _device_data: &Self::DeviceData) {}
 
+    fn pipeline_cache_get_data(
+        &self,
+        _: &Self::PipelineCacheId,
+        _: &Self::PipelineCacheData,
+    ) -> Option<Vec<u8>> {
+        None
+    }
+
     fn compute_pass_set_pipeline(
         &self,
         _pass: &mut Self::ComputePassId,

wgpu/src/backend/wgpu_core.rs

@@ -4,10 +4,10 @@ use crate::{
     BufferDescriptor, CommandEncoderDescriptor, CompilationInfo, CompilationMessage,
     CompilationMessageType, ComputePassDescriptor, ComputePipelineDescriptor,
     DownlevelCapabilities, Features, Label, Limits, LoadOp, MapMode, Operations,
-    PipelineLayoutDescriptor, RenderBundleEncoderDescriptor, RenderPipelineDescriptor,
-    SamplerDescriptor, ShaderModuleDescriptor, ShaderModuleDescriptorSpirV, ShaderSource, StoreOp,
-    SurfaceStatus, SurfaceTargetUnsafe, TextureDescriptor, TextureViewDescriptor,
-    UncapturedErrorHandler,
+    PipelineCacheDescriptor, PipelineLayoutDescriptor, RenderBundleEncoderDescriptor,
+    RenderPipelineDescriptor, SamplerDescriptor, ShaderModuleDescriptor,
+    ShaderModuleDescriptorSpirV, ShaderSource, StoreOp, SurfaceStatus, SurfaceTargetUnsafe,
+    TextureDescriptor, TextureViewDescriptor, UncapturedErrorHandler,
 };
 
 use arrayvec::ArrayVec;
@@ -519,6 +519,8 @@ impl crate::Context for ContextWgpuCore {
     type RenderPipelineData = ();
     type ComputePipelineId = wgc::id::ComputePipelineId;
     type ComputePipelineData = ();
+    type PipelineCacheId = wgc::id::PipelineCacheId;
+    type PipelineCacheData = ();
     type CommandEncoderId = wgc::id::CommandEncoderId;
     type CommandEncoderData = CommandEncoder;
     type ComputePassId = Unused;
@@ -1191,6 +1193,7 @@ impl crate::Context for ContextWgpuCore {
                 targets: Borrowed(frag.targets),
             }),
             multiview: desc.multiview,
+            cache: desc.cache.map(|c| c.id.into()),
         };
 
         let (id, error) = wgc::gfx_select!(device => self.0.device_create_render_pipeline(
@@ -1240,6 +1243,7 @@ impl crate::Context for ContextWgpuCore {
                     .compilation_options
                     .zero_initialize_workgroup_memory,
             },
+            cache: desc.cache.map(|c| c.id.into()),
         };
 
         let (id, error) = wgc::gfx_select!(device => self.0.device_create_compute_pipeline(
@@ -1267,6 +1271,37 @@ impl crate::Context for ContextWgpuCore {
         }
         (id, ())
     }
+
+    unsafe fn device_create_pipeline_cache(
+        &self,
+        device: &Self::DeviceId,
+        device_data: &Self::DeviceData,
+        desc: &PipelineCacheDescriptor<'_>,
+    ) -> (Self::PipelineCacheId, Self::PipelineCacheData) {
+        use wgc::pipeline as pipe;
+
+        let descriptor = pipe::PipelineCacheDescriptor {
+            label: desc.label.map(Borrowed),
+            data: desc.data.map(Borrowed),
+            fallback: desc.fallback,
+        };
+        let (id, error) = wgc::gfx_select!(device => self.0.device_create_pipeline_cache(
+            *device,
+            &descriptor,
+            None
+        ));
+        if let Some(cause) = error {
+            self.handle_error(
+                &device_data.error_sink,
+                cause,
+                LABEL,
+                desc.label,
+                "Device::device_create_pipeline_cache_init",
+            );
+        }
+        (id, ())
+    }
+
     fn device_create_buffer(
         &self,
         device: &Self::DeviceId,
@@ -1726,6 +1761,14 @@ impl crate::Context for ContextWgpuCore {
         wgc::gfx_select!(*pipeline => self.0.render_pipeline_drop(*pipeline))
     }
 
+    fn pipeline_cache_drop(
+        &self,
+        cache: &Self::PipelineCacheId,
+        _cache_data: &Self::PipelineCacheData,
+    ) {
+        wgc::gfx_select!(*cache => self.0.pipeline_cache_drop(*cache))
+    }
+
     fn compute_pipeline_get_bind_group_layout(
         &self,
         pipeline: &Self::ComputePipelineId,
@@ -2336,6 +2379,15 @@ impl crate::Context for ContextWgpuCore {
         wgc::gfx_select!(device => self.0.device_stop_capture(*device));
     }
 
+    fn pipeline_cache_get_data(
+        &self,
+        cache: &Self::PipelineCacheId,
+        // TODO: Used for error handling?
+        _cache_data: &Self::PipelineCacheData,
+    ) -> Option<Vec<u8>> {
+        wgc::gfx_select!(cache => self.0.pipeline_cache_get_data(*cache))
+    }
+
     fn compute_pass_set_pipeline(
         &self,
         _pass: &mut Self::ComputePassId,

wgpu/src/context.rs

@@ -11,11 +11,12 @@ use crate::{
     AnyWasmNotSendSync, BindGroupDescriptor, BindGroupLayoutDescriptor, Buffer, BufferAsyncError,
     BufferDescriptor, CommandEncoderDescriptor, CompilationInfo, ComputePassDescriptor,
     ComputePipelineDescriptor, DeviceDescriptor, Error, ErrorFilter, ImageCopyBuffer,
-    ImageCopyTexture, Maintain, MaintainResult, MapMode, PipelineLayoutDescriptor,
-    QuerySetDescriptor, RenderBundleDescriptor, RenderBundleEncoderDescriptor,
-    RenderPassDescriptor, RenderPipelineDescriptor, RequestAdapterOptions, RequestDeviceError,
-    SamplerDescriptor, ShaderModuleDescriptor, ShaderModuleDescriptorSpirV, SurfaceTargetUnsafe,
-    Texture, TextureDescriptor, TextureViewDescriptor, UncapturedErrorHandler,
+    ImageCopyTexture, Maintain, MaintainResult, MapMode, PipelineCacheDescriptor,
+    PipelineLayoutDescriptor, QuerySetDescriptor, RenderBundleDescriptor,
+    RenderBundleEncoderDescriptor, RenderPassDescriptor, RenderPipelineDescriptor,
+    RequestAdapterOptions, RequestDeviceError, SamplerDescriptor, ShaderModuleDescriptor,
+    ShaderModuleDescriptorSpirV, SurfaceTargetUnsafe, Texture, TextureDescriptor,
+    TextureViewDescriptor, UncapturedErrorHandler,
 };
 
 /// Meta trait for an id tracked by a context.
@@ -59,6 +60,8 @@ pub trait Context: Debug + WasmNotSendSync + Sized {
     type RenderPipelineData: ContextData;
     type ComputePipelineId: ContextId + WasmNotSendSync;
     type ComputePipelineData: ContextData;
+    type PipelineCacheId: ContextId + WasmNotSendSync;
+    type PipelineCacheData: ContextData;
     type CommandEncoderId: ContextId + WasmNotSendSync;
     type CommandEncoderData: ContextData;
     type ComputePassId: ContextId;
@@ -233,6 +236,12 @@ pub trait Context: Debug + WasmNotSendSync + Sized {
         device_data: &Self::DeviceData,
         desc: &ComputePipelineDescriptor<'_>,
     ) -> (Self::ComputePipelineId, Self::ComputePipelineData);
+    unsafe fn device_create_pipeline_cache(
+        &self,
+        device: &Self::DeviceId,
+        device_data: &Self::DeviceData,
+        desc: &PipelineCacheDescriptor<'_>,
+    ) -> (Self::PipelineCacheId, Self::PipelineCacheData);
     fn device_create_buffer(
         &self,
         device: &Self::DeviceId,
@@ -395,6 +404,11 @@ pub trait Context: Debug + WasmNotSendSync + Sized {
         pipeline: &Self::RenderPipelineId,
         pipeline_data: &Self::RenderPipelineData,
     );
+    fn pipeline_cache_drop(
+        &self,
+        cache: &Self::PipelineCacheId,
+        cache_data: &Self::PipelineCacheData,
+    );
 
     fn compute_pipeline_get_bind_group_layout(
         &self,
@@ -613,6 +627,12 @@ pub trait Context: Debug + WasmNotSendSync + Sized {
     fn device_start_capture(&self, device: &Self::DeviceId, device_data: &Self::DeviceData);
     fn device_stop_capture(&self, device: &Self::DeviceId, device_data: &Self::DeviceData);
 
+    fn pipeline_cache_get_data(
+        &self,
+        cache: &Self::PipelineCacheId,
+        cache_data: &Self::PipelineCacheData,
+    ) -> Option<Vec<u8>>;
+
     fn compute_pass_set_pipeline(
         &self,
         pass: &mut Self::ComputePassId,
@@ -1271,6 +1291,12 @@ pub(crate) trait DynContext: Debug + WasmNotSendSync {
         device_data: &crate::Data,
         desc: &ComputePipelineDescriptor<'_>,
     ) -> (ObjectId, Box<crate::Data>);
+    unsafe fn device_create_pipeline_cache(
+        &self,
+        device: &ObjectId,
+        device_data: &crate::Data,
+        desc: &PipelineCacheDescriptor<'_>,
+    ) -> (ObjectId, Box<crate::Data>);
     fn device_create_buffer(
         &self,
         device: &ObjectId,
@@ -1391,6 +1417,7 @@ pub(crate) trait DynContext: Debug + WasmNotSendSync {
     fn render_bundle_drop(&self, render_bundle: &ObjectId, render_bundle_data: &crate::Data);
     fn compute_pipeline_drop(&self, pipeline: &ObjectId, pipeline_data: &crate::Data);
     fn render_pipeline_drop(&self, pipeline: &ObjectId, pipeline_data: &crate::Data);
+    fn pipeline_cache_drop(&self, cache: &ObjectId, _cache_data: &crate::Data);
 
     fn compute_pipeline_get_bind_group_layout(
         &self,
@@ -1601,6 +1628,12 @@ pub(crate) trait DynContext: Debug + WasmNotSendSync {
     fn device_start_capture(&self, device: &ObjectId, data: &crate::Data);
     fn device_stop_capture(&self, device: &ObjectId, data: &crate::Data);
 
+    fn pipeline_cache_get_data(
+        &self,
+        cache: &ObjectId,
+        cache_data: &crate::Data,
+    ) -> Option<Vec<u8>>;
+
     fn compute_pass_set_pipeline(
         &self,
         pass: &mut ObjectId,
@@ -2297,6 +2330,19 @@ where
         (compute_pipeline.into(), Box::new(data) as _)
     }
 
+    unsafe fn device_create_pipeline_cache(
+        &self,
+        device: &ObjectId,
+        device_data: &crate::Data,
+        desc: &PipelineCacheDescriptor<'_>,
+    ) -> (ObjectId, Box<crate::Data>) {
+        let device = <T::DeviceId>::from(*device);
+        let device_data = downcast_ref(device_data);
+        let (pipeline_cache, data) =
+            unsafe { Context::device_create_pipeline_cache(self, &device, device_data, desc) };
+        (pipeline_cache.into(), Box::new(data) as _)
+    }
+
     fn device_create_buffer(
         &self,
         device: &ObjectId,
@@ -2621,6 +2667,12 @@ where
         Context::render_pipeline_drop(self, &pipeline, pipeline_data)
     }
 
+    fn pipeline_cache_drop(&self, cache: &ObjectId, cache_data: &crate::Data) {
+        let cache = <T::PipelineCacheId>::from(*cache);
+        let cache_data = downcast_ref(cache_data);
+        Context::pipeline_cache_drop(self, &cache, cache_data)
+    }
+
     fn compute_pipeline_get_bind_group_layout(
         &self,
         pipeline: &ObjectId,
@@ -3083,6 +3135,16 @@ where
         Context::device_stop_capture(self, &device, device_data)
     }
 
+    fn pipeline_cache_get_data(
+        &self,
+        cache: &ObjectId,
+        cache_data: &crate::Data,
+    ) -> Option<Vec<u8>> {
+        let cache = <T::PipelineCacheId>::from(*cache);
+        let cache_data = downcast_ref::<T::PipelineCacheData>(cache_data);
+        Context::pipeline_cache_get_data(self, &cache, cache_data)
+    }
+
     fn compute_pass_set_pipeline(
         &self,
         pass: &mut ObjectId,

wgpu/src/lib.rs

@@ -1105,6 +1105,100 @@ impl ComputePipeline {
     }
 }
 
+/// Handle to a pipeline cache, which is used to accelerate
+/// creating [`RenderPipeline`]s and [`ComputePipeline`]s
+/// in subsequent executions
+///
+/// This reuse is only applicable for the same or similar devices.
+/// See [`util::pipeline_cache_key`] for some details.
+///
+/// # Background
+///
+/// In most GPU drivers, shader code must be converted into a machine code
+/// which can be executed on the GPU.
+/// Generating this machine code can require a lot of computation.
+/// Pipeline caches allow this computation to be reused between executions
+/// of the program.
+/// This can be very useful for reducing program startup time.
+///
+/// Note that most desktop GPU drivers will manage their own caches,
+/// meaning that little advantage can be gained from this on those platforms.
+/// However, on some platforms, especially Android, drivers leave this to the
+/// application to implement.
+///
+/// Unfortunately, drivers do not expose whether they manage their own caches.
+/// Some reasonable policies for applications to use are:
+/// - Manage their own pipeline cache on all platforms
+/// - Only manage pipeline caches on Android
+///
+/// # Usage
+///
+/// It is valid to use this resource when creating multiple pipelines, in
+/// which case it will likely cache each of those pipelines.
+/// It is also valid to create a new cache for each pipeline.
+///
+/// This resource is most useful when the data produced from it (using
+/// [`PipelineCache::get_data`]) is persisted.
+/// Care should be taken that pipeline caches are only used for the same device,
+/// as pipeline caches from compatible devices are unlikely to provide any advantage.
+/// `util::pipeline_cache_key` can be used as a file/directory name to help ensure that.
+///
+/// It is recommended to store pipeline caches atomically. If persisting to disk,
+/// this can usually be achieved by creating a temporary file, then moving/[renaming]
+/// the temporary file over the existing cache
+///
+/// # Storage Usage
+///
+/// There is not currently an API available to reduce the size of a cache.
+/// This is due to limitations in the underlying graphics APIs used.
+/// This is especially impactful if your application is being updated, so
+/// previous caches are no longer being used.
+///
+/// One option to work around this is to regenerate the cache.
+/// That is, creating the pipelines which your program runs using
+/// with the stored cached data, then recreating the *same* pipelines
+/// using a new cache, which your application then store.
+///
+/// # Implementations
+///
+/// This resource currently only works on the following backends:
+///  - Vulkan
+///
+/// This type is unique to the Rust API of `wgpu`.
+///
+/// [renaming]: std::fs::rename
+#[derive(Debug)]
+pub struct PipelineCache {
+    context: Arc<C>,
+    id: ObjectId,
+    data: Box<Data>,
+}
+
+#[cfg(send_sync)]
+static_assertions::assert_impl_all!(PipelineCache: Send, Sync);
+
+impl PipelineCache {
+    /// Get the data associated with this pipeline cache.
+    /// The data format is an implementation detail of `wgpu`.
+    /// The only defined operation on this data setting it as the `data` field
+    /// on [`PipelineCacheDescriptor`], then to [`Device::create_pipeline_cache`].
+    ///
+    /// This function is unique to the Rust API of `wgpu`.
+    pub fn get_data(&self) -> Option<Vec<u8>> {
+        self.context
+            .pipeline_cache_get_data(&self.id, self.data.as_ref())
+    }
+}
+
+impl Drop for PipelineCache {
+    fn drop(&mut self) {
+        if !thread::panicking() {
+            self.context
+                .pipeline_cache_drop(&self.id, self.data.as_ref());
+        }
+    }
+}
+
 /// Handle to a command buffer on the GPU.
 ///
 /// A `CommandBuffer` represents a complete sequence of commands that may be submitted to a command
@@ -1832,6 +1926,8 @@ pub struct RenderPipelineDescriptor<'a> {
     /// If the pipeline will be used with a multiview render pass, this indicates how many array
     /// layers the attachments will have.
     pub multiview: Option<NonZeroU32>,
+    /// The pipeline cache to use when creating this pipeline.
+    pub cache: Option<&'a PipelineCache>,
 }
 #[cfg(send_sync)]
 static_assertions::assert_impl_all!(RenderPipelineDescriptor<'_>: Send, Sync);
@@ -1929,10 +2025,38 @@ pub struct ComputePipelineDescriptor<'a> {
     ///
     /// This implements `Default`, and for most users can be set to `Default::default()`
     pub compilation_options: PipelineCompilationOptions<'a>,
+    /// The pipeline cache to use when creating this pipeline.
+    pub cache: Option<&'a PipelineCache>,
 }
 #[cfg(send_sync)]
 static_assertions::assert_impl_all!(ComputePipelineDescriptor<'_>: Send, Sync);
 
+/// Describes a pipeline cache, which allows reusing compilation work
+/// between program runs.
+///
+/// For use with [`Device::create_pipeline_cache`]
+///
+/// This type is unique to the Rust API of `wgpu`.
+#[derive(Clone, Debug)]
+pub struct PipelineCacheDescriptor<'a> {
+    /// Debug label of the pipeline cache. This might show up in some logs from `wgpu`
+    pub label: Label<'a>,
+    /// The data used to initialise the cache initialise
+    ///
+    /// # Safety
+    ///
+    /// This data must have been provided from a previous call to
+    /// [`PipelineCache::get_data`], if not `None`
+    pub data: Option<&'a [u8]>,
+    /// Whether to create a cache without data when the provided data
+    /// is invalid.
+    ///
+    /// Recommended to set to true
+    pub fallback: bool,
+}
+#[cfg(send_sync)]
+static_assertions::assert_impl_all!(PipelineCacheDescriptor<'_>: Send, Sync);
+
 pub use wgt::ImageCopyBuffer as ImageCopyBufferBase;
 /// View of a buffer which can be used to copy to/from a texture.
 ///
@@ -3080,6 +3204,62 @@ impl Device {
     pub fn make_invalid(&self) {
         DynContext::device_make_invalid(&*self.context, &self.id, self.data.as_ref())
     }
+
+    /// Create a [`PipelineCache`] with initial data
+    ///
+    /// This can be passed to [`Device::create_compute_pipeline`]
+    /// and [`Device::create_render_pipeline`] to either accelerate these
+    /// or add the cache results from those.
+    ///
+    /// # Safety
+    ///
+    /// If the `data` field of `desc` is set, it must have previously been returned from a call
+    /// to [`PipelineCache::get_data`][^saving]. This `data` will only be used if it came
+    /// from an adapter with the same [`util::pipeline_cache_key`].
+    /// This *is* compatible across wgpu versions, as any data format change will
+    /// be accounted for.
+    ///
+    /// It is *not* supported to bring caches from previous direct uses of backend APIs
+    /// into this method.
+    ///
+    /// # Errors
+    ///
+    /// Returns an error value if:
+    ///  * the [`PIPELINE_CACHE`](wgt::Features::PIPELINE_CACHE) feature is not enabled
+    ///  * this device is invalid; or
+    ///  * the device is out of memory
+    ///
+    /// This method also returns an error value if:
+    ///  * The `fallback` field on `desc` is false; and
+    ///  * the `data` provided would not be used[^data_not_used]
+    ///
+    /// If an error value is used in subsequent calls, default caching will be used.
+    ///
+    /// [^saving]: We do recognise that saving this data to disk means this condition
+    /// is impossible to fully prove. Consider the risks for your own application in this case.
+    ///
+    /// [^data_not_used]: This data may be not used if: the data was produced by a prior
+    /// version of wgpu; or was created for an incompatible adapter, or there was a GPU driver
+    /// update. In some cases, the data might not be used and a real value is returned,
+    /// this is left to the discretion of GPU drivers.
+    pub unsafe fn create_pipeline_cache(
+        &self,
+        desc: &PipelineCacheDescriptor<'_>,
+    ) -> PipelineCache {
+        let (id, data) = unsafe {
+            DynContext::device_create_pipeline_cache(
+                &*self.context,
+                &self.id,
+                self.data.as_ref(),
+                desc,
+            )
+        };
+        PipelineCache {
+            context: Arc::clone(&self.context),
+            id,
+            data,
+        }
+    }
 }
 
 impl Drop for Device {

wgpu/src/util/mod.rs

@@ -140,3 +140,52 @@ impl std::ops::Deref for DownloadBuffer {
         self.1.slice()
     }
 }
+
+/// A recommended key for storing [`PipelineCache`]s for the adapter
+/// associated with the given [`AdapterInfo`](wgt::AdapterInfo)
+/// This key will define a class of adapters for which the same cache
+/// might be valid.
+///
+/// If this returns `None`, the adapter doesn't support [`PipelineCache`].
+/// This may be because the API doesn't support application managed caches
+/// (such as browser WebGPU), or that `wgpu` hasn't implemented it for
+/// that API yet.
+///
+/// This key could be used as a filename, as seen in the example below.
+///
+/// # Examples
+///
+/// ``` no_run
+/// # use std::path::PathBuf;
+/// # let adapter_info = todo!();
+/// let cache_dir: PathBuf = PathBuf::new();
+/// let filename = wgpu::util::pipeline_cache_key(&adapter_info);
+/// if let Some(filename) = filename {
+///     let cache_file = cache_dir.join(&filename);
+///     let cache_data = std::fs::read(&cache_file);
+///     let pipeline_cache: wgpu::PipelineCache = todo!("Use data (if present) to create a pipeline cache");
+///
+///     let data = pipeline_cache.get_data();
+///     if let Some(data) = data {
+///         let temp_file = cache_file.with_extension("temp");
+///         std::fs::write(&temp_file, &data)?;
+///         std::fs::rename(&temp_file, &cache_file)?;
+///     }
+/// }
+/// # Ok::<(), std::io::Error>(())
+/// ```
+///
+/// [`PipelineCache`]: super::PipelineCache
+pub fn pipeline_cache_key(adapter_info: &wgt::AdapterInfo) -> Option<String> {
+    match adapter_info.backend {
+        wgt::Backend::Vulkan => Some(format!(
+            // The vendor/device should uniquely define a driver
+            // We/the driver will also later validate that the vendor/device and driver
+            // version match, which may lead to clearing an outdated
+            // cache for the same device.
+            "wgpu_pipeline_cache_vulkan_{}_{}",
+            adapter_info.vendor, adapter_info.device
+        )),
+        _ => None,
+    }
+}