ext/webgpu/device.rs - main - githubmirror/deno

ext/webgpu/device.rs
1 065 строк30 KB
Leo Kettmeir
chore: update webgpu (#34319)
08 июн 2026, 23:29
08 июн 2026, 23:292b313e2
// Copyright 2018-2026 the Deno authors. MIT license.

use std::borrow::Cow;
use std::cell::RefCell;
use std::num::NonZeroU64;
use std::rc::Rc;

use deno_core::GarbageCollected;
use deno_core::cppgc::SameObject;
use deno_core::cppgc::make_cppgc_object;
use deno_core::op2;
use deno_core::v8;
use deno_core::webidl::WebIdlInterfaceConverter;
use deno_error::JsErrorBox;
use wgpu_core::binding_model::BindingResource;
use wgpu_core::pipeline::ProgrammableStageDescriptor;
use wgpu_types::BindingType;

use super::bind_group::GPUBindGroup;
use super::bind_group::GPUBindingResource;
use super::bind_group_layout::GPUBindGroupLayout;
use super::buffer::GPUBuffer;
use super::compute_pipeline::GPUComputePipeline;
use super::pipeline_layout::GPUPipelineLayout;
use super::sampler::GPUSampler;
use super::shader::GPUShaderModule;
use super::texture::GPUTexture;
use crate::Instance;
use crate::adapter::GPUAdapterInfo;
use crate::adapter::GPUSupportedFeatures;
use crate::adapter::GPUSupportedLimits;
use crate::command_encoder::GPUCommandEncoder;
use crate::error::GPUError;
use crate::error::GPUGenericError;
use crate::error::GPUPipelineErrorReason;
use crate::error::fmt_err;
use crate::error::make_pipeline_error;
use crate::query_set::GPUQuerySet;
use crate::render_bundle::GPURenderBundleEncoder;
use crate::render_pipeline::GPURenderPipeline;
use crate::shader::GPUCompilationInfo;

/// External memory associated with device and queue, to encourage V8 to garbage
/// collect devices promptly. This seems to be particularly important when
/// running CTS tests under `webgpu:api,validation,capability_checks,limits,*`
/// on DX12 in wgpu CI, where any smaller power of two results in OOM errors.
pub(crate) const DEVICE_EXTERNAL_MEMORY_SIZE: i64 = 1 << 24; // 16 MB

pub struct GPUDevice {
  pub instance: Instance,
  pub id: wgpu_core::id::DeviceId,
  pub adapter: wgpu_core::id::AdapterId,
  pub queue: wgpu_core::id::QueueId,

  pub label: String,

  pub features: SameObject<GPUSupportedFeatures>,
  pub limits: SameObject<GPUSupportedLimits>,
  pub adapter_info: Rc<SameObject<GPUAdapterInfo>>,

  pub queue_obj: v8::Global<v8::Object>,

  pub error_handler: super::error::ErrorHandler,
  pub lost_promise: v8::Global<v8::Promise>,

  // Weak reference to the JS object so we can attach a finalizer.
  pub(crate) weak: std::sync::OnceLock<v8::Weak<v8::Object>>,
  pub has_active_capture: RefCell<bool>,
}

impl Drop for GPUDevice {
  fn drop(&mut self) {
    self.instance.device_drop(self.id);
  }
}

impl WebIdlInterfaceConverter for GPUDevice {
  const NAME: &'static str = "GPUDevice";
}

// SAFETY: we're sure this can be GCed
unsafe impl GarbageCollected for GPUDevice {
  fn trace(&self, _visitor: &mut deno_core::v8::cppgc::Visitor) {}

  fn get_name(&self) -> &'static std::ffi::CStr {
    c"GPUDevice"
  }
}

// EventTarget is extended in JS
#[op2]
impl GPUDevice {
  #[constructor]
  #[cppgc]
  fn constructor(_: bool) -> Result<GPUDevice, GPUGenericError> {
    Err(GPUGenericError::InvalidConstructor)
  }

  #[getter]
  #[string]
  fn label(&self) -> String {
    self.label.clone()
  }
  #[setter]
  #[string]
  fn label(&self, #[webidl] _label: String) {
    // TODO(@crowlKats): no-op, needs wpgu to implement changing the label
  }

  #[getter]
  fn features(
    &self,
    scope: &mut v8::PinScope<'_, '_>,
  ) -> v8::Global<v8::Object> {
    self.features.get(scope, |scope| {
      let features = self.instance.device_features(self.id);
      GPUSupportedFeatures::new(scope, features)
    })
  }

  #[getter]
  fn limits(&self, scope: &mut v8::PinScope<'_, '_>) -> v8::Global<v8::Object> {
    self.limits.get(scope, |_| {
      let limits = self.instance.device_limits(self.id);
      GPUSupportedLimits(limits)
    })
  }

  #[getter]
  fn adapter_info(
    &self,
    scope: &mut v8::PinScope<'_, '_>,
  ) -> v8::Global<v8::Object> {
    self.adapter_info.get(scope, |_| {
      let info = self.instance.adapter_get_info(self.adapter);

      GPUAdapterInfo { info }
    })
  }

  #[getter]
  fn queue(&self) -> v8::Global<v8::Object> {
    self.queue_obj.clone()
  }

  #[fast]
  #[undefined]
  fn destroy(&self) {
    self.instance.device_destroy(self.id);
    self
      .error_handler
      .push_error(Some(GPUError::Lost(GPUDeviceLostReason::Destroyed)));
  }

  #[required(1)]
  #[cppgc]
  fn create_buffer(
    &self,
    #[webidl] descriptor: super::buffer::GPUBufferDescriptor,
  ) -> Result<GPUBuffer, JsErrorBox> {
    // wgpu-core would also check this, but it needs to be reported via a JS
    // error, not a validation error. (WebGPU specifies this check on the
    // content timeline.)
    if descriptor.mapped_at_creation
      && !descriptor
        .size
        .is_multiple_of(wgpu_types::COPY_BUFFER_ALIGNMENT)
    {
      return Err(JsErrorBox::range_error(format!(
        "The size of a buffer that is mapped at creation must be a multiple of {}",
        wgpu_types::COPY_BUFFER_ALIGNMENT,
      )));
    }

    // Validation of the usage needs to happen on the device timeline, so
    // don't raise an error immediately if it isn't valid. wgpu will
    // reject `BufferUsages::empty()`.
    let usage = wgpu_types::BufferUsages::from_bits(descriptor.usage)
      .unwrap_or(wgpu_types::BufferUsages::empty());

    let wgpu_descriptor = wgpu_core::resource::BufferDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      size: descriptor.size,
      usage,
      mapped_at_creation: descriptor.mapped_at_creation,
    };

    let (id, err) =
      self
        .instance
        .device_create_buffer(self.id, &wgpu_descriptor, None);

    self.error_handler.push_error(err);

    Ok(GPUBuffer {
      instance: self.instance.clone(),
      error_handler: self.error_handler.clone(),
      id,
      device: self.id,
      label: descriptor.label,
      size: descriptor.size,
      usage: descriptor.usage,
      map_state: RefCell::new(if descriptor.mapped_at_creation {
        "mapped"
      } else {
        "unmapped"
      }),
      map_mode: RefCell::new(if descriptor.mapped_at_creation {
        Some(wgpu_core::device::HostMap::Write)
      } else {
        None
      }),
      mapped_js_buffers: RefCell::new(vec![]),
    })
  }

  #[required(1)]
  #[cppgc]
  fn create_texture(
    &self,
    #[webidl] descriptor: super::texture::GPUTextureDescriptor,
  ) -> Result<GPUTexture, JsErrorBox> {
    let wgpu_descriptor = wgpu_core::resource::TextureDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      size: descriptor.size.into(),
      mip_level_count: descriptor.mip_level_count,
      sample_count: descriptor.sample_count,
      dimension: descriptor.dimension.clone().into(),
      format: descriptor.format.clone().into(),
      usage: descriptor.usage.into(),
      view_formats: descriptor
        .view_formats
        .into_iter()
        .map(Into::into)
        .collect(),
    };

    let (id, err) =
      self
        .instance
        .device_create_texture(self.id, &wgpu_descriptor, None);

    self.error_handler.push_error(err);

    Ok(GPUTexture {
      instance: self.instance.clone(),
      error_handler: self.error_handler.clone(),
      id,
      device_id: self.id,
      queue_id: self.queue,
      default_view_id: Default::default(),
      label: descriptor.label,
      size: wgpu_descriptor.size,
      mip_level_count: wgpu_descriptor.mip_level_count,
      sample_count: wgpu_descriptor.sample_count,
      dimension: descriptor.dimension,
      format: descriptor.format,
      usage: descriptor.usage,
    })
  }

  #[cppgc]
  fn create_sampler(
    &self,
    #[webidl] descriptor: super::sampler::GPUSamplerDescriptor,
  ) -> Result<GPUSampler, JsErrorBox> {
    let wgpu_descriptor = wgpu_core::resource::SamplerDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      address_modes: [
        descriptor.address_mode_u.into(),
        descriptor.address_mode_v.into(),
        descriptor.address_mode_w.into(),
      ],
      mag_filter: descriptor.mag_filter.into(),
      min_filter: descriptor.min_filter.into(),
      mipmap_filter: descriptor.mipmap_filter.into(),
      lod_min_clamp: descriptor.lod_min_clamp,
      lod_max_clamp: descriptor.lod_max_clamp,
      compare: descriptor.compare.map(Into::into),
      anisotropy_clamp: descriptor.max_anisotropy,
      border_color: None,
    };

    let (id, err) =
      self
        .instance
        .device_create_sampler(self.id, &wgpu_descriptor, None);

    self.error_handler.push_error(err);

    Ok(GPUSampler {
      instance: self.instance.clone(),
      id,
      label: descriptor.label,
    })
  }

  #[required(1)]
  #[cppgc]
  fn create_bind_group_layout(
    &self,
    #[webidl]
    descriptor: super::bind_group_layout::GPUBindGroupLayoutDescriptor,
  ) -> Result<GPUBindGroupLayout, JsErrorBox> {
    let mut entries = Vec::with_capacity(descriptor.entries.len());

    for entry in descriptor.entries {
      let n_entries = [
        entry.buffer.is_some(),
        entry.sampler.is_some(),
        entry.texture.is_some(),
        entry.storage_texture.is_some(),
        entry.external_texture.is_some(),
      ]
      .into_iter()
      .filter(|t| *t)
      .count();

      if n_entries != 1 {
        return Err(JsErrorBox::type_error(
          "Only one of 'buffer', 'sampler', 'texture' and 'storageTexture' may be specified",
        ));
      }

      let ty = if let Some(buffer) = entry.buffer {
        BindingType::Buffer {
          ty: buffer.r#type.into(),
          has_dynamic_offset: buffer.has_dynamic_offset,
          min_binding_size: NonZeroU64::new(buffer.min_binding_size),
        }
      } else if let Some(sampler) = entry.sampler {
        BindingType::Sampler(sampler.r#type.into())
      } else if let Some(texture) = entry.texture {
        BindingType::Texture {
          sample_type: texture.sample_type.into(),
          view_dimension: texture.view_dimension.into(),
          multisampled: texture.multisampled,
        }
      } else if let Some(storage_texture) = entry.storage_texture {
        BindingType::StorageTexture {
          access: storage_texture.access.into(),
          format: storage_texture.format.into(),
          view_dimension: storage_texture.view_dimension.into(),
        }
      } else if entry.external_texture.is_some() {
        BindingType::ExternalTexture
      } else {
        unreachable!()
      };

      entries.push(wgpu_types::BindGroupLayoutEntry {
        binding: entry.binding,
        visibility: entry.visibility.into(),
        ty,
        count: None, // native-only
      });
    }

    let wgpu_descriptor = wgpu_core::binding_model::BindGroupLayoutDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      entries: Cow::Owned(entries),
    };

    let (id, err) = self.instance.device_create_bind_group_layout(
      self.id,
      &wgpu_descriptor,
      None,
    );

    self.error_handler.push_error(err);

    Ok(GPUBindGroupLayout {
      instance: self.instance.clone(),
      id,
      label: descriptor.label,
    })
  }

  #[required(1)]
  #[cppgc]
  fn create_pipeline_layout(
    &self,
    #[webidl] descriptor: super::pipeline_layout::GPUPipelineLayoutDescriptor,
  ) -> GPUPipelineLayout {
    let bind_group_layouts = descriptor
      .bind_group_layouts
      .into_iter()
      .map(|bind_group_layout| {
        bind_group_layout
          .into_option()
          .map(|bind_group_layout| bind_group_layout.id)
      })
      .collect();

    let wgpu_descriptor = wgpu_core::binding_model::PipelineLayoutDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      bind_group_layouts: Cow::Owned(bind_group_layouts),
      immediate_size: 0,
    };

    let (id, err) = self.instance.device_create_pipeline_layout(
      self.id,
      &wgpu_descriptor,
      None,
    );

    self.error_handler.push_error(err);

    GPUPipelineLayout {
      instance: self.instance.clone(),
      id,
      label: descriptor.label,
    }
  }

  #[required(1)]
  #[cppgc]
  fn create_bind_group(
    &self,
    #[webidl] descriptor: super::bind_group::GPUBindGroupDescriptor,
  ) -> GPUBindGroup {
    let entries = descriptor
      .entries
      .into_iter()
      .map(|entry| wgpu_core::binding_model::BindGroupEntry {
        binding: entry.binding,
        resource: match entry.resource {
          GPUBindingResource::Sampler(sampler) => {
            BindingResource::Sampler(sampler.id)
          }
          GPUBindingResource::Texture(texture) => {
            BindingResource::TextureView(texture.default_view_id())
          }
          GPUBindingResource::TextureView(texture_view) => {
            BindingResource::TextureView(texture_view.id)
          }
          GPUBindingResource::Buffer(buffer) => {
            BindingResource::Buffer(wgpu_core::binding_model::BufferBinding {
              buffer: buffer.id,
              offset: 0,
              size: Some(buffer.size),
            })
          }
          GPUBindingResource::BufferBinding(buffer_binding) => {
            BindingResource::Buffer(wgpu_core::binding_model::BufferBinding {
              buffer: buffer_binding.buffer.id,
              offset: buffer_binding.offset,
              size: buffer_binding.size,
            })
          }
          GPUBindingResource::ExternalTexture(external_texture) => {
            BindingResource::ExternalTexture(external_texture.id)
          }
        },
      })
      .collect::<Vec<_>>();

    let wgpu_descriptor = wgpu_core::binding_model::BindGroupDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      layout: descriptor.layout.id,
      entries: Cow::Owned(entries),
    };

    let (id, err) =
      self
        .instance
        .device_create_bind_group(self.id, &wgpu_descriptor, None);

    self.error_handler.push_error(err);

    GPUBindGroup {
      instance: self.instance.clone(),
      id,
      label: descriptor.label,
    }
  }

  #[required(1)]
  #[cppgc]
  fn create_shader_module(
    &self,
    scope: &mut v8::PinScope<'_, '_>,
    #[webidl] descriptor: super::shader::GPUShaderModuleDescriptor,
  ) -> GPUShaderModule {
    let wgpu_descriptor = wgpu_core::pipeline::ShaderModuleDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      runtime_checks: wgpu_types::ShaderRuntimeChecks::default(),
    };

    let (id, err) = self.instance.device_create_shader_module(
      self.id,
      &wgpu_descriptor,
      wgpu_core::pipeline::ShaderModuleSource::Wgsl(Cow::Borrowed(
        &descriptor.code,
      )),
      None,
    );

    let compilation_info =
      GPUCompilationInfo::new(scope, err.iter(), &descriptor.code);
    let compilation_info = make_cppgc_object(scope, compilation_info);
    let compilation_info = v8::Global::new(scope, compilation_info);
    self.error_handler.push_error(err);

    GPUShaderModule {
      instance: self.instance.clone(),
      id,
      label: descriptor.label,
      compilation_info,
    }
  }

  #[required(1)]
  #[cppgc]
  fn create_compute_pipeline(
    &self,
    #[webidl] descriptor: super::compute_pipeline::GPUComputePipelineDescriptor,
  ) -> GPUComputePipeline {
    let (pipeline, err) = self.new_compute_pipeline(descriptor);
    self.error_handler.push_error(err);
    pipeline
  }

  #[required(1)]
  #[cppgc]
  fn create_render_pipeline(
    &self,
    #[webidl] descriptor: super::render_pipeline::GPURenderPipelineDescriptor,
  ) -> GPURenderPipeline {
    let (pipeline, err) = self.new_render_pipeline(descriptor);
    self.error_handler.push_error(err);
    pipeline
  }

  #[async_method(fake)]
  #[required(1)]
  fn create_compute_pipeline_async(
    &self,
    scope: &mut v8::PinScope<'_, '_>,
    #[webidl] descriptor: super::compute_pipeline::GPUComputePipelineDescriptor,
  ) -> v8::Global<v8::Promise> {
    let resolver = v8::PromiseResolver::new(scope).unwrap();
    let promise = resolver.get_promise(scope);

    let (pipeline, err) = self.new_compute_pipeline(descriptor);
    if let Some(err) = err {
      let err = make_pipeline_error(
        scope,
        GPUPipelineErrorReason::Validation,
        &fmt_err(&err),
      );
      resolver.reject(scope, err.into());
    } else {
      let val = make_cppgc_object(scope, pipeline).into();
      resolver.resolve(scope, val);
    }
    v8::Global::new(scope, promise)
  }

  #[async_method(fake)]
  #[required(1)]
  fn create_render_pipeline_async(
    &self,
    scope: &mut v8::PinScope<'_, '_>,
    #[webidl] descriptor: super::render_pipeline::GPURenderPipelineDescriptor,
  ) -> v8::Global<v8::Promise> {
    let resolver = v8::PromiseResolver::new(scope).unwrap();
    let promise = resolver.get_promise(scope);

    let (pipeline, err) = self.new_render_pipeline(descriptor);
    if let Some(err) = err {
      let err = make_pipeline_error(
        scope,
        GPUPipelineErrorReason::Validation,
        &fmt_err(&err),
      );
      resolver.reject(scope, err.into());
    } else {
      let val = make_cppgc_object(scope, pipeline).into();
      resolver.resolve(scope, val);
    }
    v8::Global::new(scope, promise)
  }

  fn create_command_encoder<'a>(
    &self,
    scope: &mut v8::PinScope<'a, '_>,
    #[webidl] descriptor: Option<
      super::command_encoder::GPUCommandEncoderDescriptor,
    >,
  ) -> v8::Local<'a, v8::Object> {
    // Metal imposes a limit on the number of outstanding command buffers.
    // Attempting to create another command buffer after reaching that limit
    // will block, which can result in a deadlock if GC is required to
    // recover old command buffers. To encourage V8 to garbage collect
    // command buffers before that happens, we associate some external
    // memory with each command buffer.
    #[cfg(target_vendor = "apple")]
    const EXTERNAL_MEMORY_AMOUNT: i64 = 1 << 16;

    let label = descriptor.map(|d| d.label).unwrap_or_default();
    let wgpu_descriptor = wgpu_types::CommandEncoderDescriptor {
      label: Some(Cow::Owned(label.clone())),
    };

    #[cfg(target_vendor = "apple")]
    scope.adjust_amount_of_external_allocated_memory(EXTERNAL_MEMORY_AMOUNT);

    let (id, err) = self.instance.device_create_command_encoder(
      self.id,
      &wgpu_descriptor,
      None,
    );

    self.error_handler.push_error(err);

    let encoder = GPUCommandEncoder {
      instance: self.instance.clone(),
      error_handler: self.error_handler.clone(),
      id,
      label,
      #[cfg(target_vendor = "apple")]
      weak: std::sync::OnceLock::new(),
    };

    let obj = make_cppgc_object(scope, encoder);

    #[cfg(target_vendor = "apple")]
    {
      let finalizer = v8::Weak::with_finalizer(
        scope,
        obj,
        Box::new(|isolate: &mut v8::Isolate| {
          isolate.adjust_amount_of_external_allocated_memory(
            -EXTERNAL_MEMORY_AMOUNT,
          );
        }),
      );
      deno_core::cppgc::try_unwrap_cppgc_object::<GPUCommandEncoder>(
        scope,
        obj.into(),
      )
      .unwrap()
      .weak
      .set(finalizer)
      .unwrap();
    }

    obj
  }

  #[required(1)]
  #[cppgc]
  fn create_render_bundle_encoder(
    &self,
    #[webidl]
    descriptor: super::render_bundle::GPURenderBundleEncoderDescriptor,
  ) -> GPURenderBundleEncoder {
    let wgpu_descriptor = wgpu_core::command::RenderBundleEncoderDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      color_formats: Cow::Owned(
        descriptor
          .color_formats
          .into_iter()
          .map(|format| format.into_option().map(Into::into))
          .collect::<Vec<_>>(),
      ),
      depth_stencil: descriptor.depth_stencil_format.map(|format| {
        wgpu_types::RenderBundleDepthStencil {
          format: format.into(),
          depth_read_only: descriptor.depth_read_only,
          stencil_read_only: descriptor.stencil_read_only,
        }
      }),
      sample_count: descriptor.sample_count,
      multiview: None,
    };

    let res =
      wgpu_core::command::RenderBundleEncoder::new(&wgpu_descriptor, self.id);
    let (encoder, err) = match res {
      Ok(encoder) => (encoder, None),
      Err(e) => (
        wgpu_core::command::RenderBundleEncoder::dummy(self.id),
        Some(e),
      ),
    };

    self.error_handler.push_error(err);

    GPURenderBundleEncoder {
      instance: self.instance.clone(),
      error_handler: self.error_handler.clone(),
      encoder: RefCell::new(Some(encoder)),
      label: descriptor.label,
    }
  }

  #[required(1)]
  #[cppgc]
  fn create_query_set(
    &self,
    #[webidl] descriptor: crate::query_set::GPUQuerySetDescriptor,
  ) -> GPUQuerySet {
    let wgpu_descriptor = wgpu_core::resource::QuerySetDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      ty: descriptor.r#type.clone().into(),
      count: descriptor.count,
    };

    let (id, err) =
      self
        .instance
        .device_create_query_set(self.id, &wgpu_descriptor, None);

    self.error_handler.push_error(err);

    GPUQuerySet {
      instance: self.instance.clone(),
      id,
      r#type: descriptor.r#type,
      count: descriptor.count,
      label: descriptor.label,
    }
  }

  #[getter]
  fn lost(&self) -> v8::Global<v8::Promise> {
    self.lost_promise.clone()
  }

  #[required(1)]
  #[undefined]
  fn push_error_scope(&self, #[webidl] filter: super::error::GPUErrorFilter) {
    self
      .error_handler
      .scopes
      .lock()
      .unwrap()
      .push((filter, None));
  }

  #[async_method(fake)]
  fn pop_error_scope(
    &self,
    scope: &mut v8::PinScope<'_, '_>,
  ) -> Result<v8::Global<v8::Value>, JsErrorBox> {
    if self.error_handler.is_lost.get().is_some() {
      let val = v8::null(scope).cast::<v8::Value>();
      return Ok(v8::Global::new(scope, val));
    }

    let Some((_, error)) = self.error_handler.scopes.lock().unwrap().pop()
    else {
      return Err(JsErrorBox::new(
        "DOMExceptionOperationError",
        "There are no error scopes on the error scope stack",
      ));
    };

    let val = if let Some(err) = error {
      deno_core::error::to_v8_error(scope, &err)
    } else {
      v8::null(scope).into()
    };

    Ok(v8::Global::new(scope, val))
  }
}

impl GPUDevice {
  fn new_compute_pipeline(
    &self,
    descriptor: super::compute_pipeline::GPUComputePipelineDescriptor,
  ) -> (
    GPUComputePipeline,
    Option<wgpu_core::pipeline::CreateComputePipelineError>,
  ) {
    let wgpu_descriptor = wgpu_core::pipeline::ComputePipelineDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      layout: descriptor.layout.into(),
      stage: ProgrammableStageDescriptor {
        module: descriptor.compute.module.id,
        entry_point: descriptor.compute.entry_point.map(Into::into),
        constants: descriptor.compute.constants.into_iter().collect(),
        zero_initialize_workgroup_memory: true,
      },
      cache: None,
    };

    let (id, err) = self.instance.device_create_compute_pipeline(
      self.id,
      &wgpu_descriptor,
      None,
    );

    (
      GPUComputePipeline {
        instance: self.instance.clone(),
        error_handler: self.error_handler.clone(),
        id,
        label: descriptor.label.clone(),
      },
      err,
    )
  }

  fn new_render_pipeline(
    &self,
    descriptor: super::render_pipeline::GPURenderPipelineDescriptor,
  ) -> (
    GPURenderPipeline,
    Option<wgpu_core::pipeline::CreateRenderPipelineError>,
  ) {
    let vertex = wgpu_core::pipeline::VertexState {
      stage: ProgrammableStageDescriptor {
        module: descriptor.vertex.module.id,
        entry_point: descriptor.vertex.entry_point.map(Into::into),
        constants: descriptor.vertex.constants.into_iter().collect(),
        zero_initialize_workgroup_memory: true,
      },
      buffers: Cow::Owned(
        descriptor
          .vertex
          .buffers
          .into_iter()
          .map(|b| {
            b.into_option().map_or_else(
              wgpu_core::pipeline::VertexBufferLayout::default,
              |layout| wgpu_core::pipeline::VertexBufferLayout {
                array_stride: layout.array_stride,
                step_mode: layout.step_mode.into(),
                attributes: Cow::Owned(
                  layout
                    .attributes
                    .into_iter()
                    .map(|attr| wgpu_types::VertexAttribute {
                      format: attr.format.into(),
                      offset: attr.offset,
                      shader_location: attr.shader_location,
                    })
                    .collect(),
                ),
              },
            )
          })
          .collect(),
      ),
    };

    let primitive = wgpu_types::PrimitiveState {
      topology: descriptor.primitive.topology.into(),
      strip_index_format: descriptor
        .primitive
        .strip_index_format
        .map(Into::into),
      front_face: descriptor.primitive.front_face.into(),
      cull_mode: descriptor.primitive.cull_mode.into(),
      unclipped_depth: descriptor.primitive.unclipped_depth,
      polygon_mode: Default::default(),
      conservative: false,
    };

    let depth_stencil = descriptor.depth_stencil.map(|depth_stencil| {
      let front = wgpu_types::StencilFaceState {
        compare: depth_stencil.stencil_front.compare.into(),
        fail_op: depth_stencil.stencil_front.fail_op.into(),
        depth_fail_op: depth_stencil.stencil_front.depth_fail_op.into(),
        pass_op: depth_stencil.stencil_front.pass_op.into(),
      };
      let back = wgpu_types::StencilFaceState {
        compare: depth_stencil.stencil_back.compare.into(),
        fail_op: depth_stencil.stencil_back.fail_op.into(),
        depth_fail_op: depth_stencil.stencil_back.depth_fail_op.into(),
        pass_op: depth_stencil.stencil_back.pass_op.into(),
      };

      wgpu_types::DepthStencilState {
        format: depth_stencil.format.into(),
        depth_write_enabled: depth_stencil.depth_write_enabled,
        depth_compare: depth_stencil.depth_compare.map(Into::into),
        stencil: wgpu_types::StencilState {
          front,
          back,
          read_mask: depth_stencil.stencil_read_mask,
          write_mask: depth_stencil.stencil_write_mask,
        },
        bias: wgpu_types::DepthBiasState {
          constant: depth_stencil.depth_bias,
          slope_scale: depth_stencil.depth_bias_slope_scale,
          clamp: depth_stencil.depth_bias_clamp,
        },
      }
    });

    let multisample = wgpu_types::MultisampleState {
      count: descriptor.multisample.count,
      mask: descriptor.multisample.mask as u64,
      alpha_to_coverage_enabled: descriptor
        .multisample
        .alpha_to_coverage_enabled,
    };

    let fragment =
      descriptor
        .fragment
        .map(|fragment| wgpu_core::pipeline::FragmentState {
          stage: ProgrammableStageDescriptor {
            module: fragment.module.id,
            entry_point: fragment.entry_point.map(Into::into),
            constants: fragment.constants.into_iter().collect(),
            zero_initialize_workgroup_memory: true,
          },
          targets: Cow::Owned(
            fragment
              .targets
              .into_iter()
              .map(|target| {
                target.into_option().map(|target| {
                  wgpu_types::ColorTargetState {
                    format: target.format.into(),
                    blend: target.blend.map(|blend| wgpu_types::BlendState {
                      color: wgpu_types::BlendComponent {
                        src_factor: blend.color.src_factor.into(),
                        dst_factor: blend.color.dst_factor.into(),
                        operation: blend.color.operation.into(),
                      },
                      alpha: wgpu_types::BlendComponent {
                        src_factor: blend.alpha.src_factor.into(),
                        dst_factor: blend.alpha.dst_factor.into(),
                        operation: blend.alpha.operation.into(),
                      },
                    }),
                    write_mask: target.write_mask.into(),
                  }
                })
              })
              .collect(),
          ),
        });

    let wgpu_descriptor = wgpu_core::pipeline::RenderPipelineDescriptor {
      label: crate::transform_label(descriptor.label.clone()),
      layout: descriptor.layout.into(),
      vertex,
      primitive,
      depth_stencil,
      multisample,
      fragment,
      cache: None,
      multiview_mask: None,
    };

    let (id, err) = self.instance.device_create_render_pipeline(
      self.id,
      &wgpu_descriptor,
      None,
    );

    (
      GPURenderPipeline {
        instance: self.instance.clone(),
        error_handler: self.error_handler.clone(),
        id,
        label: descriptor.label,
      },
      err,
    )
  }
}

#[derive(Clone, Debug, Default, Hash, Eq, PartialEq)]
pub enum GPUDeviceLostReason {
  #[default]
  Unknown,
  Destroyed,
}

impl From<wgpu_types::DeviceLostReason> for GPUDeviceLostReason {
  fn from(value: wgpu_types::DeviceLostReason) -> Self {
    match value {
      wgpu_types::DeviceLostReason::Unknown => Self::Unknown,
      wgpu_types::DeviceLostReason::Destroyed => Self::Destroyed,
    }
  }
}

#[derive(Default)]
pub struct GPUDeviceLostInfo {
  pub reason: GPUDeviceLostReason,
}

// SAFETY: we're sure this can be GCed
unsafe impl GarbageCollected for GPUDeviceLostInfo {
  fn trace(&self, _visitor: &mut deno_core::v8::cppgc::Visitor) {}

  fn get_name(&self) -> &'static std::ffi::CStr {
    c"GPUDeviceLostInfo"
  }
}

#[op2]
impl GPUDeviceLostInfo {
  #[constructor]
  #[cppgc]
  fn constructor(_: bool) -> Result<GPUDeviceLostInfo, GPUGenericError> {
    Err(GPUGenericError::InvalidConstructor)
  }

  #[getter]
  #[string]
  fn reason(&self) -> &'static str {
    use GPUDeviceLostReason::*;
    match self.reason {
      Unknown => "unknown",
      Destroyed => "destroyed",
    }
  }

  #[getter]
  #[string]
  fn message(&self) -> &'static str {
    "device was lost"
  }
}

#[op2(fast)]
pub fn op_webgpu_device_start_capture(
  #[cppgc] device: &GPUDevice,
) -> Result<(), JsErrorBox> {
  if *device.has_active_capture.borrow() {
    return Err(JsErrorBox::type_error("capture already started"));
  }

  // safety: active check is above, other safety concerns are related to the debugger itself
  unsafe {
    device
      .instance
      .device_start_graphics_debugger_capture(device.id);
  }

  *device.has_active_capture.borrow_mut() = true;

  Ok(())
}

#[op2(fast)]
pub fn op_webgpu_device_stop_capture(
  #[cppgc] device: &GPUDevice,
) -> Result<(), JsErrorBox> {
  if !*device.has_active_capture.borrow() {
    return Err(JsErrorBox::type_error("No capture active"));
  }

  // safety: active check is above, other safety concerns are related to the debugger itself
  unsafe {
    device
      .instance
      .device_stop_graphics_debugger_capture(device.id);
  }

  *device.has_active_capture.borrow_mut() = false;

  Ok(())
}