Quelle  command.rs   Sprache: unbekannt

use super::{conv, AsNative, TimestampQuerySupport};
use crate::CommandEncoder as _;
use std::{borrow::Cow, mem::size_of, ops::Range};

// has to match `Temp::binding_sizes`
const WORD_SIZE: usize = 4;

impl Default for super::CommandState {
    fn default() -> Self {
        Self {
            blit: None,
            render: None,
            compute: None,
            raw_primitive_type: metal::MTLPrimitiveType::Point,
            index: None,
            raw_wg_size: metal::MTLSize::new(0, 0, 0),
            stage_infos: Default::default(),
            storage_buffer_length_map: Default::default(),
            vertex_buffer_size_map: Default::default(),
            work_group_memory_sizes: Vec::new(),
            push_constants: Vec::new(),
            pending_timer_queries: Vec::new(),
        }
    }
}

impl super::CommandEncoder {
    fn enter_blit(&mut self) -> &metal::BlitCommandEncoderRef {
        if self.state.blit.is_none() {
            debug_assert!(self.state.render.is_none() && self.state.compute.is_none());
            let cmd_buf = self.raw_cmd_buf.as_ref().unwrap();

            // Take care of pending timer queries.
            // If we can't use `sample_counters_in_buffer` we have to create a dummy blit encoder!
            //
            // There is a known bug in Metal where blit encoders won't write timestamps if they don't have a blit operation.
            // See https://github.com/gpuweb/gpuweb/issues/2046#issuecomment-1205793680 & https://source.chromium.org/chromium/chromium/src/+/006c4eb70c96229834bbaf271290f40418144cd3:third_party/dawn/src/dawn/native/metal/BackendMTL.mm;l=350
            //
            // To make things worse:
            // * what counts as a blit operation is a bit unclear, experimenting seemed to indicate that resolve_counters doesn't count.
            // * in some cases (when?) using `set_start_of_encoder_sample_index` doesn't work, so we have to use `set_end_of_encoder_sample_index` instead
            //
            // All this means that pretty much the only *reliable* thing as of writing is to:
            // * create a dummy blit encoder using set_end_of_encoder_sample_index
            // * do a dummy write that is known to be not optimized out.
            // * close the encoder since we used set_end_of_encoder_sample_index and don't want to get any extra stuff in there.
            // * create another encoder for whatever we actually had in mind.
            let supports_sample_counters_in_buffer = self
                .shared
                .private_caps
                .timestamp_query_support
                .contains(TimestampQuerySupport::ON_BLIT_ENCODER);

            if !self.state.pending_timer_queries.is_empty() && !supports_sample_counters_in_buffer {
                objc::rc::autoreleasepool(|| {
                    let descriptor = metal::BlitPassDescriptor::new();
                    let mut last_query = None;
                    for (i, (set, index)) in self.state.pending_timer_queries.drain(..).enumerate()
                    {
                        let sba_descriptor = descriptor
                            .sample_buffer_attachments()
                            .object_at(i as _)
                            .unwrap();
                        sba_descriptor
                            .set_sample_buffer(set.counter_sample_buffer.as_ref().unwrap());

                        // Here be dragons:
                        // As mentioned above, for some reasons using the start of the encoder won't yield any results sometimes!
                        sba_descriptor
                            .set_start_of_encoder_sample_index(metal::COUNTER_DONT_SAMPLE);
                        sba_descriptor.set_end_of_encoder_sample_index(index as _);

                        last_query = Some((set, index));
                    }
                    let encoder = cmd_buf.blit_command_encoder_with_descriptor(descriptor);

                    // As explained above, we need to do some write:
                    // Conveniently, we have a buffer with every query set, that we can use for this for a dummy write,
                    // since we know that it is going to be overwritten again on timer resolve and HAL doesn't define its state before that.
                    let raw_range = metal::NSRange {
                        location: last_query.as_ref().unwrap().1 as u64 * crate::QUERY_SIZE,
                        length: 1,
                    };
                    encoder.fill_buffer(
                        &last_query.as_ref().unwrap().0.raw_buffer,
                        raw_range,
                        255, // Don't write 0, so it's easier to identify if something went wrong.
                    );

                    encoder.end_encoding();
                });
            }

            objc::rc::autoreleasepool(|| {
                self.state.blit = Some(cmd_buf.new_blit_command_encoder().to_owned());
            });

            let encoder = self.state.blit.as_ref().unwrap();

            // UNTESTED:
            // If the above described issue with empty blit encoder applies to `sample_counters_in_buffer` as well, we should use the same workaround instead!
            for (set, index) in self.state.pending_timer_queries.drain(..) {
                debug_assert!(supports_sample_counters_in_buffer);
                encoder.sample_counters_in_buffer(
                    set.counter_sample_buffer.as_ref().unwrap(),
                    index as _,
                    true,
                )
            }
        }
        self.state.blit.as_ref().unwrap()
    }

    pub(super) fn leave_blit(&mut self) {
        if let Some(encoder) = self.state.blit.take() {
            encoder.end_encoding();
        }
    }

    fn active_encoder(&mut self) -> Option<&metal::CommandEncoderRef> {
        if let Some(ref encoder) = self.state.render {
            Some(encoder)
        } else if let Some(ref encoder) = self.state.compute {
            Some(encoder)
        } else if let Some(ref encoder) = self.state.blit {
            Some(encoder)
        } else {
            None
        }
    }

    fn begin_pass(&mut self) {
        self.state.reset();
        self.leave_blit();
    }
}

impl super::CommandState {
    fn reset(&mut self) {
        self.storage_buffer_length_map.clear();
        self.vertex_buffer_size_map.clear();
        self.stage_infos.vs.clear();
        self.stage_infos.fs.clear();
        self.stage_infos.cs.clear();
        self.work_group_memory_sizes.clear();
        self.push_constants.clear();
    }

    fn make_sizes_buffer_update<'a>(
        &self,
        stage: naga::ShaderStage,
        result_sizes: &'a mut Vec<u32>,
    ) -> Option<(u32, &'a [u32])> {
        let stage_info = &self.stage_infos[stage];
        let slot = stage_info.sizes_slot?;

        result_sizes.clear();
        result_sizes.extend(stage_info.sized_bindings.iter().map(|br| {
            self.storage_buffer_length_map
                .get(br)
                .map(|size| u32::try_from(size.get()).unwrap_or(u32::MAX))
                .unwrap_or_default()
        }));

        // Extend with the sizes of the mapped vertex buffers, in the order
        // they were added to the map.
        result_sizes.extend(stage_info.vertex_buffer_mappings.iter().map(|vbm| {
            self.vertex_buffer_size_map
                .get(&(vbm.id as u64))
                .map(|size| u32::try_from(size.get()).unwrap_or(u32::MAX))
                .unwrap_or_default()
        }));

        if !result_sizes.is_empty() {
            Some((slot as _, result_sizes))
        } else {
            None
        }
    }
}

impl crate::CommandEncoder for super::CommandEncoder {
    type A = super::Api;

    unsafe fn begin_encoding(&mut self, label: crate::Label) -> Result<(), crate::DeviceError> {
        let queue = &self.raw_queue.lock();
        let retain_references = self.shared.settings.retain_command_buffer_references;
        let raw = objc::rc::autoreleasepool(move || {
            let cmd_buf_ref = if retain_references {
                queue.new_command_buffer()
            } else {
                queue.new_command_buffer_with_unretained_references()
            };
            if let Some(label) = label {
                cmd_buf_ref.set_label(label);
            }
            cmd_buf_ref.to_owned()
        });

        self.raw_cmd_buf = Some(raw);

        Ok(())
    }

    unsafe fn discard_encoding(&mut self) {
        self.leave_blit();
        // when discarding, we don't have a guarantee that
        // everything is in a good state, so check carefully
        if let Some(encoder) = self.state.render.take() {
            encoder.end_encoding();
        }
        if let Some(encoder) = self.state.compute.take() {
            encoder.end_encoding();
        }
        self.raw_cmd_buf = None;
    }

    unsafe fn end_encoding(&mut self) -> Result<super::CommandBuffer, crate::DeviceError> {
        // Handle pending timer query if any.
        if !self.state.pending_timer_queries.is_empty() {
            self.leave_blit();
            self.enter_blit();
        }

        self.leave_blit();
        debug_assert!(self.state.render.is_none());
        debug_assert!(self.state.compute.is_none());
        debug_assert!(self.state.pending_timer_queries.is_empty());

        Ok(super::CommandBuffer {
            raw: self.raw_cmd_buf.take().unwrap(),
        })
    }

    unsafe fn reset_all<I>(&mut self, _cmd_bufs: I)
    where
        I: Iterator<Item = super::CommandBuffer>,
    {
        //do nothing
    }

    unsafe fn transition_buffers<'a, T>(&mut self, _barriers: T)
    where
        T: Iterator<Item = crate::BufferBarrier<'a, super::Buffer>>,
    {
    }

    unsafe fn transition_textures<'a, T>(&mut self, _barriers: T)
    where
        T: Iterator<Item = crate::TextureBarrier<'a, super::Texture>>,
    {
    }

    unsafe fn clear_buffer(&mut self, buffer: &super::Buffer, range: crate::MemoryRange) {
        let encoder = self.enter_blit();
        encoder.fill_buffer(&buffer.raw, conv::map_range(&range), 0);
    }

    unsafe fn copy_buffer_to_buffer<T>(
        &mut self,
        src: &super::Buffer,
        dst: &super::Buffer,
        regions: T,
    ) where
        T: Iterator<Item = crate::BufferCopy>,
    {
        let encoder = self.enter_blit();
        for copy in regions {
            encoder.copy_from_buffer(
                &src.raw,
                copy.src_offset,
                &dst.raw,
                copy.dst_offset,
                copy.size.get(),
            );
        }
    }

    unsafe fn copy_texture_to_texture<T>(
        &mut self,
        src: &super::Texture,
        _src_usage: crate::TextureUses,
        dst: &super::Texture,
        regions: T,
    ) where
        T: Iterator<Item = crate::TextureCopy>,
    {
        let dst_texture = if src.format != dst.format {
            let raw_format = self.shared.private_caps.map_format(src.format);
            Cow::Owned(objc::rc::autoreleasepool(|| {
                dst.raw.new_texture_view(raw_format)
            }))
        } else {
            Cow::Borrowed(&dst.raw)
        };
        let encoder = self.enter_blit();
        for copy in regions {
            let src_origin = conv::map_origin(©.src_base.origin);
            let dst_origin = conv::map_origin(©.dst_base.origin);
            // no clamping is done: Metal expects physical sizes here
            let extent = conv::map_copy_extent(©.size);
            encoder.copy_from_texture(
                &src.raw,
                copy.src_base.array_layer as u64,
                copy.src_base.mip_level as u64,
                src_origin,
                extent,
                &dst_texture,
                copy.dst_base.array_layer as u64,
                copy.dst_base.mip_level as u64,
                dst_origin,
            );
        }
    }

    unsafe fn copy_buffer_to_texture<T>(
        &mut self,
        src: &super::Buffer,
        dst: &super::Texture,
        regions: T,
    ) where
        T: Iterator<Item = crate::BufferTextureCopy>,
    {
        let encoder = self.enter_blit();
        for copy in regions {
            let dst_origin = conv::map_origin(©.texture_base.origin);
            // Metal expects buffer-texture copies in virtual sizes
            let extent = copy
                .texture_base
                .max_copy_size(&dst.copy_size)
                .min(©.size);
            let bytes_per_row = copy.buffer_layout.bytes_per_row.unwrap_or(0) as u64;
            let image_byte_stride = if extent.depth > 1 {
                copy.buffer_layout
                    .rows_per_image
                    .map_or(0, |v| v as u64 * bytes_per_row)
            } else {
                // Don't pass a stride when updating a single layer, otherwise metal validation
                // fails when updating a subset of the image due to the stride being larger than
                // the amount of data to copy.
                0
            };
            encoder.copy_from_buffer_to_texture(
                &src.raw,
                copy.buffer_layout.offset,
                bytes_per_row,
                image_byte_stride,
                conv::map_copy_extent(&extent),
                &dst.raw,
                copy.texture_base.array_layer as u64,
                copy.texture_base.mip_level as u64,
                dst_origin,
                conv::get_blit_option(dst.format, copy.texture_base.aspect),
            );
        }
    }

    unsafe fn copy_texture_to_buffer<T>(
        &mut self,
        src: &super::Texture,
        _src_usage: crate::TextureUses,
        dst: &super::Buffer,
        regions: T,
    ) where
        T: Iterator<Item = crate::BufferTextureCopy>,
    {
        let encoder = self.enter_blit();
        for copy in regions {
            let src_origin = conv::map_origin(©.texture_base.origin);
            // Metal expects texture-buffer copies in virtual sizes
            let extent = copy
                .texture_base
                .max_copy_size(&src.copy_size)
                .min(©.size);
            let bytes_per_row = copy.buffer_layout.bytes_per_row.unwrap_or(0) as u64;
            let bytes_per_image = copy
                .buffer_layout
                .rows_per_image
                .map_or(0, |v| v as u64 * bytes_per_row);
            encoder.copy_from_texture_to_buffer(
                &src.raw,
                copy.texture_base.array_layer as u64,
                copy.texture_base.mip_level as u64,
                src_origin,
                conv::map_copy_extent(&extent),
                &dst.raw,
                copy.buffer_layout.offset,
                bytes_per_row,
                bytes_per_image,
                conv::get_blit_option(src.format, copy.texture_base.aspect),
            );
        }
    }

    unsafe fn begin_query(&mut self, set: &super::QuerySet, index: u32) {
        match set.ty {
            wgt::QueryType::Occlusion => {
                self.state
                    .render
                    .as_ref()
                    .unwrap()
                    .set_visibility_result_mode(
                        metal::MTLVisibilityResultMode::Boolean,
                        index as u64 * crate::QUERY_SIZE,
                    );
            }
            _ => {}
        }
    }
    unsafe fn end_query(&mut self, set: &super::QuerySet, _index: u32) {
        match set.ty {
            wgt::QueryType::Occlusion => {
                self.state
                    .render
                    .as_ref()
                    .unwrap()
                    .set_visibility_result_mode(metal::MTLVisibilityResultMode::Disabled, 0);
            }
            _ => {}
        }
    }
    unsafe fn write_timestamp(&mut self, set: &super::QuerySet, index: u32) {
        let support = self.shared.private_caps.timestamp_query_support;
        debug_assert!(
            support.contains(TimestampQuerySupport::STAGE_BOUNDARIES),
            "Timestamp queries are not supported"
        );
        let sample_buffer = set.counter_sample_buffer.as_ref().unwrap();
        let with_barrier = true;

        // Try to use an existing encoder for timestamp query if possible.
        // This works only if it's supported for the active encoder.
        if let (true, Some(encoder)) = (
            support.contains(TimestampQuerySupport::ON_BLIT_ENCODER),
            self.state.blit.as_ref(),
        ) {
            encoder.sample_counters_in_buffer(sample_buffer, index as _, with_barrier);
        } else if let (true, Some(encoder)) = (
            support.contains(TimestampQuerySupport::ON_RENDER_ENCODER),
            self.state.render.as_ref(),
        ) {
            encoder.sample_counters_in_buffer(sample_buffer, index as _, with_barrier);
        } else if let (true, Some(encoder)) = (
            support.contains(TimestampQuerySupport::ON_COMPUTE_ENCODER),
            self.state.compute.as_ref(),
        ) {
            encoder.sample_counters_in_buffer(sample_buffer, index as _, with_barrier);
        } else {
            // If we're here it means we either have no encoder open, or it's not supported to sample within them.
            // If this happens with render/compute open, this is an invalid usage!
            debug_assert!(self.state.render.is_none() && self.state.compute.is_none());

            // But otherwise it means we'll put defer this to the next created encoder.
            self.state.pending_timer_queries.push((set.clone(), index));

            // Ensure we didn't already have a blit open.
            self.leave_blit();
        };
    }

    unsafe fn reset_queries(&mut self, set: &super::QuerySet, range: Range<u32>) {
        let encoder = self.enter_blit();
        let raw_range = metal::NSRange {
            location: range.start as u64 * crate::QUERY_SIZE,
            length: (range.end - range.start) as u64 * crate::QUERY_SIZE,
        };
        encoder.fill_buffer(&set.raw_buffer, raw_range, 0);
    }

    unsafe fn copy_query_results(
        &mut self,
        set: &super::QuerySet,
        range: Range<u32>,
        buffer: &super::Buffer,
        offset: wgt::BufferAddress,
        _: wgt::BufferSize, // Metal doesn't support queries that are bigger than a single element are not supported
    ) {
        let encoder = self.enter_blit();
        match set.ty {
            wgt::QueryType::Occlusion => {
                let size = (range.end - range.start) as u64 * crate::QUERY_SIZE;
                encoder.copy_from_buffer(
                    &set.raw_buffer,
                    range.start as u64 * crate::QUERY_SIZE,
                    &buffer.raw,
                    offset,
                    size,
                );
            }
            wgt::QueryType::Timestamp => {
                encoder.resolve_counters(
                    set.counter_sample_buffer.as_ref().unwrap(),
                    metal::NSRange::new(range.start as u64, (range.end - range.start) as u64),
                    &buffer.raw,
                    offset,
                );
            }
            wgt::QueryType::PipelineStatistics(_) => todo!(),
        }
    }

    // render

    unsafe fn begin_render_pass(
        &mut self,
        desc: &crate::RenderPassDescriptor<super::QuerySet, super::TextureView>,
    ) {
        self.begin_pass();
        self.state.index = None;

        assert!(self.state.blit.is_none());
        assert!(self.state.compute.is_none());
        assert!(self.state.render.is_none());

        objc::rc::autoreleasepool(|| {
            let descriptor = metal::RenderPassDescriptor::new();

            for (i, at) in desc.color_attachments.iter().enumerate() {
                if let Some(at) = at.as_ref() {
                    let at_descriptor = descriptor.color_attachments().object_at(i as u64).unwrap();
                    at_descriptor.set_texture(Some(&at.target.view.raw));
                    if let Some(ref resolve) = at.resolve_target {
                        //Note: the selection of levels and slices is already handled by `TextureView`
                        at_descriptor.set_resolve_texture(Some(&resolve.view.raw));
                    }
                    let load_action = if at.ops.contains(crate::AttachmentOps::LOAD) {
                        metal::MTLLoadAction::Load
                    } else {
                        at_descriptor.set_clear_color(conv::map_clear_color(&at.clear_value));
                        metal::MTLLoadAction::Clear
                    };
                    let store_action = conv::map_store_action(
                        at.ops.contains(crate::AttachmentOps::STORE),
                        at.resolve_target.is_some(),
                    );
                    at_descriptor.set_load_action(load_action);
                    at_descriptor.set_store_action(store_action);
                }
            }

            if let Some(ref at) = desc.depth_stencil_attachment {
                if at.target.view.aspects.contains(crate::FormatAspects::DEPTH) {
                    let at_descriptor = descriptor.depth_attachment().unwrap();
                    at_descriptor.set_texture(Some(&at.target.view.raw));

                    let load_action = if at.depth_ops.contains(crate::AttachmentOps::LOAD) {
                        metal::MTLLoadAction::Load
                    } else {
                        at_descriptor.set_clear_depth(at.clear_value.0 as f64);
                        metal::MTLLoadAction::Clear
                    };
                    let store_action = if at.depth_ops.contains(crate::AttachmentOps::STORE) {
                        metal::MTLStoreAction::Store
                    } else {
                        metal::MTLStoreAction::DontCare
                    };
                    at_descriptor.set_load_action(load_action);
                    at_descriptor.set_store_action(store_action);
                }
                if at
                    .target
                    .view
                    .aspects
                    .contains(crate::FormatAspects::STENCIL)
                {
                    let at_descriptor = descriptor.stencil_attachment().unwrap();
                    at_descriptor.set_texture(Some(&at.target.view.raw));

                    let load_action = if at.stencil_ops.contains(crate::AttachmentOps::LOAD) {
                        metal::MTLLoadAction::Load
                    } else {
                        at_descriptor.set_clear_stencil(at.clear_value.1);
                        metal::MTLLoadAction::Clear
                    };
                    let store_action = if at.stencil_ops.contains(crate::AttachmentOps::STORE) {
                        metal::MTLStoreAction::Store
                    } else {
                        metal::MTLStoreAction::DontCare
                    };
                    at_descriptor.set_load_action(load_action);
                    at_descriptor.set_store_action(store_action);
                }
            }

            let mut sba_index = 0;
            let mut next_sba_descriptor = || {
                let sba_descriptor = descriptor
                    .sample_buffer_attachments()
                    .object_at(sba_index)
                    .unwrap();

                sba_descriptor.set_end_of_vertex_sample_index(metal::COUNTER_DONT_SAMPLE);
                sba_descriptor.set_start_of_fragment_sample_index(metal::COUNTER_DONT_SAMPLE);

                sba_index += 1;
                sba_descriptor
            };

            for (set, index) in self.state.pending_timer_queries.drain(..) {
                let sba_descriptor = next_sba_descriptor();
                sba_descriptor.set_sample_buffer(set.counter_sample_buffer.as_ref().unwrap());
                sba_descriptor.set_start_of_vertex_sample_index(index as _);
                sba_descriptor.set_end_of_fragment_sample_index(metal::COUNTER_DONT_SAMPLE);
            }

            if let Some(ref timestamp_writes) = desc.timestamp_writes {
                let sba_descriptor = next_sba_descriptor();
                sba_descriptor.set_sample_buffer(
                    timestamp_writes
                        .query_set
                        .counter_sample_buffer
                        .as_ref()
                        .unwrap(),
                );

                sba_descriptor.set_start_of_vertex_sample_index(
                    timestamp_writes
                        .beginning_of_pass_write_index
                        .map_or(metal::COUNTER_DONT_SAMPLE, |i| i as _),
                );
                sba_descriptor.set_end_of_fragment_sample_index(
                    timestamp_writes
                        .end_of_pass_write_index
                        .map_or(metal::COUNTER_DONT_SAMPLE, |i| i as _),
                );
            }

            if let Some(occlusion_query_set) = desc.occlusion_query_set {
                descriptor
                    .set_visibility_result_buffer(Some(occlusion_query_set.raw_buffer.as_ref()))
            }

            let raw = self.raw_cmd_buf.as_ref().unwrap();
            let encoder = raw.new_render_command_encoder(descriptor);
            if let Some(label) = desc.label {
                encoder.set_label(label);
            }
            self.state.render = Some(encoder.to_owned());
        });
    }

    unsafe fn end_render_pass(&mut self) {
        self.state.render.take().unwrap().end_encoding();
    }

    unsafe fn set_bind_group(
        &mut self,
        layout: &super::PipelineLayout,
        group_index: u32,
        group: &super::BindGroup,
        dynamic_offsets: &[wgt::DynamicOffset],
    ) {
        let bg_info = &layout.bind_group_infos[group_index as usize];

        if let Some(ref encoder) = self.state.render {
            let mut changes_sizes_buffer = false;
            for index in 0..group.counters.vs.buffers {
                let buf = &group.buffers[index as usize];
                let mut offset = buf.offset;
                if let Some(dyn_index) = buf.dynamic_index {
                    offset += dynamic_offsets[dyn_index as usize] as wgt::BufferAddress;
                }
                encoder.set_vertex_buffer(
                    (bg_info.base_resource_indices.vs.buffers + index) as u64,
                    Some(buf.ptr.as_native()),
                    offset,
                );
                if let Some(size) = buf.binding_size {
                    let br = naga::ResourceBinding {
                        group: group_index,
                        binding: buf.binding_location,
                    };
                    self.state.storage_buffer_length_map.insert(br, size);
                    changes_sizes_buffer = true;
                }
            }
            if changes_sizes_buffer {
                if let Some((index, sizes)) = self.state.make_sizes_buffer_update(
                    naga::ShaderStage::Vertex,
                    &mut self.temp.binding_sizes,
                ) {
                    encoder.set_vertex_bytes(
                        index as _,
                        (sizes.len() * WORD_SIZE) as u64,
                        sizes.as_ptr().cast(),
                    );
                }
            }

            changes_sizes_buffer = false;
            for index in 0..group.counters.fs.buffers {
                let buf = &group.buffers[(group.counters.vs.buffers + index) as usize];
                let mut offset = buf.offset;
                if let Some(dyn_index) = buf.dynamic_index {
                    offset += dynamic_offsets[dyn_index as usize] as wgt::BufferAddress;
                }
                encoder.set_fragment_buffer(
                    (bg_info.base_resource_indices.fs.buffers + index) as u64,
                    Some(buf.ptr.as_native()),
                    offset,
                );
                if let Some(size) = buf.binding_size {
                    let br = naga::ResourceBinding {
                        group: group_index,
                        binding: buf.binding_location,
                    };
                    self.state.storage_buffer_length_map.insert(br, size);
                    changes_sizes_buffer = true;
                }
            }
            if changes_sizes_buffer {
                if let Some((index, sizes)) = self.state.make_sizes_buffer_update(
                    naga::ShaderStage::Fragment,
                    &mut self.temp.binding_sizes,
                ) {
                    encoder.set_fragment_bytes(
                        index as _,
                        (sizes.len() * WORD_SIZE) as u64,
                        sizes.as_ptr().cast(),
                    );
                }
            }

            for index in 0..group.counters.vs.samplers {
                let res = group.samplers[index as usize];
                encoder.set_vertex_sampler_state(
                    (bg_info.base_resource_indices.vs.samplers + index) as u64,
                    Some(res.as_native()),
                );
            }
            for index in 0..group.counters.fs.samplers {
                let res = group.samplers[(group.counters.vs.samplers + index) as usize];
                encoder.set_fragment_sampler_state(
                    (bg_info.base_resource_indices.fs.samplers + index) as u64,
                    Some(res.as_native()),
                );
            }

            for index in 0..group.counters.vs.textures {
                let res = group.textures[index as usize];
                encoder.set_vertex_texture(
                    (bg_info.base_resource_indices.vs.textures + index) as u64,
                    Some(res.as_native()),
                );
            }
            for index in 0..group.counters.fs.textures {
                let res = group.textures[(group.counters.vs.textures + index) as usize];
                encoder.set_fragment_texture(
                    (bg_info.base_resource_indices.fs.textures + index) as u64,
                    Some(res.as_native()),
                );
            }

            // Call useResource on all textures and buffers used indirectly so they are alive
            for (resource, use_info) in group.resources_to_use.iter() {
                encoder.use_resource_at(resource.as_native(), use_info.uses, use_info.stages);
            }
        }

        if let Some(ref encoder) = self.state.compute {
            let index_base = super::ResourceData {
                buffers: group.counters.vs.buffers + group.counters.fs.buffers,
                samplers: group.counters.vs.samplers + group.counters.fs.samplers,
                textures: group.counters.vs.textures + group.counters.fs.textures,
            };

            let mut changes_sizes_buffer = false;
            for index in 0..group.counters.cs.buffers {
                let buf = &group.buffers[(index_base.buffers + index) as usize];
                let mut offset = buf.offset;
                if let Some(dyn_index) = buf.dynamic_index {
                    offset += dynamic_offsets[dyn_index as usize] as wgt::BufferAddress;
                }
                encoder.set_buffer(
                    (bg_info.base_resource_indices.cs.buffers + index) as u64,
                    Some(buf.ptr.as_native()),
                    offset,
                );
                if let Some(size) = buf.binding_size {
                    let br = naga::ResourceBinding {
                        group: group_index,
                        binding: buf.binding_location,
                    };
                    self.state.storage_buffer_length_map.insert(br, size);
                    changes_sizes_buffer = true;
                }
            }
            if changes_sizes_buffer {
                if let Some((index, sizes)) = self.state.make_sizes_buffer_update(
                    naga::ShaderStage::Compute,
                    &mut self.temp.binding_sizes,
                ) {
                    encoder.set_bytes(
                        index as _,
                        (sizes.len() * WORD_SIZE) as u64,
                        sizes.as_ptr().cast(),
                    );
                }
            }

            for index in 0..group.counters.cs.samplers {
                let res = group.samplers[(index_base.samplers + index) as usize];
                encoder.set_sampler_state(
                    (bg_info.base_resource_indices.cs.samplers + index) as u64,
                    Some(res.as_native()),
                );
            }
            for index in 0..group.counters.cs.textures {
                let res = group.textures[(index_base.textures + index) as usize];
                encoder.set_texture(
                    (bg_info.base_resource_indices.cs.textures + index) as u64,
                    Some(res.as_native()),
                );
            }

            // Call useResource on all textures and buffers used indirectly so they are alive
            for (resource, use_info) in group.resources_to_use.iter() {
                if !use_info.visible_in_compute {
                    continue;
                }
                encoder.use_resource(resource.as_native(), use_info.uses);
            }
        }
    }

    unsafe fn set_push_constants(
        &mut self,
        layout: &super::PipelineLayout,
        stages: wgt::ShaderStages,
        offset_bytes: u32,
        data: &[u32],
    ) {
        let state_pc = &mut self.state.push_constants;
        if state_pc.len() < layout.total_push_constants as usize {
            state_pc.resize(layout.total_push_constants as usize, 0);
        }
        debug_assert_eq!(offset_bytes as usize % WORD_SIZE, 0);

        let offset_words = offset_bytes as usize / WORD_SIZE;
        state_pc[offset_words..offset_words + data.len()].copy_from_slice(data);

        if stages.contains(wgt::ShaderStages::COMPUTE) {
            self.state.compute.as_ref().unwrap().set_bytes(
                layout.push_constants_infos.cs.unwrap().buffer_index as _,
                (layout.total_push_constants as usize * WORD_SIZE) as _,
                state_pc.as_ptr().cast(),
            )
        }
        if stages.contains(wgt::ShaderStages::VERTEX) {
            self.state.render.as_ref().unwrap().set_vertex_bytes(
                layout.push_constants_infos.vs.unwrap().buffer_index as _,
                (layout.total_push_constants as usize * WORD_SIZE) as _,
                state_pc.as_ptr().cast(),
            )
        }
        if stages.contains(wgt::ShaderStages::FRAGMENT) {
            self.state.render.as_ref().unwrap().set_fragment_bytes(
                layout.push_constants_infos.fs.unwrap().buffer_index as _,
                (layout.total_push_constants as usize * WORD_SIZE) as _,
                state_pc.as_ptr().cast(),
            )
        }
    }

    unsafe fn insert_debug_marker(&mut self, label: &str) {
        if let Some(encoder) = self.active_encoder() {
            encoder.insert_debug_signpost(label);
        }
    }
    unsafe fn begin_debug_marker(&mut self, group_label: &str) {
        if let Some(encoder) = self.active_encoder() {
            encoder.push_debug_group(group_label);
        } else if let Some(ref buf) = self.raw_cmd_buf {
            buf.push_debug_group(group_label);
        }
    }
    unsafe fn end_debug_marker(&mut self) {
        if let Some(encoder) = self.active_encoder() {
            encoder.pop_debug_group();
        } else if let Some(ref buf) = self.raw_cmd_buf {
            buf.pop_debug_group();
        }
    }

    unsafe fn set_render_pipeline(&mut self, pipeline: &super::RenderPipeline) {
        self.state.raw_primitive_type = pipeline.raw_primitive_type;
        self.state.stage_infos.vs.assign_from(&pipeline.vs_info);
        match pipeline.fs_info {
            Some(ref info) => self.state.stage_infos.fs.assign_from(info),
            None => self.state.stage_infos.fs.clear(),
        }

        let encoder = self.state.render.as_ref().unwrap();
        encoder.set_render_pipeline_state(&pipeline.raw);
        encoder.set_front_facing_winding(pipeline.raw_front_winding);
        encoder.set_cull_mode(pipeline.raw_cull_mode);
        encoder.set_triangle_fill_mode(pipeline.raw_triangle_fill_mode);
        if let Some(depth_clip) = pipeline.raw_depth_clip_mode {
            encoder.set_depth_clip_mode(depth_clip);
        }
        if let Some((ref state, bias)) = pipeline.depth_stencil {
            encoder.set_depth_stencil_state(state);
            encoder.set_depth_bias(bias.constant as f32, bias.slope_scale, bias.clamp);
        }

        {
            if let Some((index, sizes)) = self
                .state
                .make_sizes_buffer_update(naga::ShaderStage::Vertex, &mut self.temp.binding_sizes)
            {
                encoder.set_vertex_bytes(
                    index as _,
                    (sizes.len() * WORD_SIZE) as u64,
                    sizes.as_ptr().cast(),
                );
            }
        }
        if pipeline.fs_lib.is_some() {
            if let Some((index, sizes)) = self
                .state
                .make_sizes_buffer_update(naga::ShaderStage::Fragment, &mut self.temp.binding_sizes)
            {
                encoder.set_fragment_bytes(
                    index as _,
                    (sizes.len() * WORD_SIZE) as u64,
                    sizes.as_ptr().cast(),
                );
            }
        }
    }

    unsafe fn set_index_buffer<'a>(
        &mut self,
        binding: crate::BufferBinding<'a, super::Buffer>,
        format: wgt::IndexFormat,
    ) {
        let (stride, raw_type) = match format {
            wgt::IndexFormat::Uint16 => (2, metal::MTLIndexType::UInt16),
            wgt::IndexFormat::Uint32 => (4, metal::MTLIndexType::UInt32),
        };
        self.state.index = Some(super::IndexState {
            buffer_ptr: AsNative::from(binding.buffer.raw.as_ref()),
            offset: binding.offset,
            stride,
            raw_type,
        });
    }

    unsafe fn set_vertex_buffer<'a>(
        &mut self,
        index: u32,
        binding: crate::BufferBinding<'a, super::Buffer>,
    ) {
        let buffer_index = self.shared.private_caps.max_vertex_buffers as u64 - 1 - index as u64;
        let encoder = self.state.render.as_ref().unwrap();
        encoder.set_vertex_buffer(buffer_index, Some(&binding.buffer.raw), binding.offset);

        let buffer_size = binding.resolve_size();
        if buffer_size > 0 {
            self.state.vertex_buffer_size_map.insert(
                buffer_index,
                std::num::NonZeroU64::new(buffer_size).unwrap(),
            );
        } else {
            self.state.vertex_buffer_size_map.remove(&buffer_index);
        }

        if let Some((index, sizes)) = self
            .state
            .make_sizes_buffer_update(naga::ShaderStage::Vertex, &mut self.temp.binding_sizes)
        {
            encoder.set_vertex_bytes(
                index as _,
                (sizes.len() * WORD_SIZE) as u64,
                sizes.as_ptr().cast(),
            );
        }
    }

    unsafe fn set_viewport(&mut self, rect: &crate::Rect<f32>, depth_range: Range<f32>) {
        let zfar = if self.shared.disabilities.broken_viewport_near_depth {
            depth_range.end - depth_range.start
        } else {
            depth_range.end
        };
        let encoder = self.state.render.as_ref().unwrap();
        encoder.set_viewport(metal::MTLViewport {
            originX: rect.x as _,
            originY: rect.y as _,
            width: rect.w as _,
            height: rect.h as _,
            znear: depth_range.start as _,
            zfar: zfar as _,
        });
    }
    unsafe fn set_scissor_rect(&mut self, rect: &crate::Rect<u32>) {
        //TODO: support empty scissors by modifying the viewport
        let scissor = metal::MTLScissorRect {
            x: rect.x as _,
            y: rect.y as _,
            width: rect.w as _,
            height: rect.h as _,
        };
        let encoder = self.state.render.as_ref().unwrap();
        encoder.set_scissor_rect(scissor);
    }
    unsafe fn set_stencil_reference(&mut self, value: u32) {
        let encoder = self.state.render.as_ref().unwrap();
        encoder.set_stencil_front_back_reference_value(value, value);
    }
    unsafe fn set_blend_constants(&mut self, color: &[f32; 4]) {
        let encoder = self.state.render.as_ref().unwrap();
        encoder.set_blend_color(color[0], color[1], color[2], color[3]);
    }

    unsafe fn draw(
        &mut self,
        first_vertex: u32,
        vertex_count: u32,
        first_instance: u32,
        instance_count: u32,
    ) {
        let encoder = self.state.render.as_ref().unwrap();
        if first_instance != 0 {
            encoder.draw_primitives_instanced_base_instance(
                self.state.raw_primitive_type,
                first_vertex as _,
                vertex_count as _,
                instance_count as _,
                first_instance as _,
            );
        } else if instance_count != 1 {
            encoder.draw_primitives_instanced(
                self.state.raw_primitive_type,
                first_vertex as _,
                vertex_count as _,
                instance_count as _,
            );
        } else {
            encoder.draw_primitives(
                self.state.raw_primitive_type,
                first_vertex as _,
                vertex_count as _,
            );
        }
    }

    unsafe fn draw_indexed(
        &mut self,
        first_index: u32,
        index_count: u32,
        base_vertex: i32,
        first_instance: u32,
        instance_count: u32,
    ) {
        let encoder = self.state.render.as_ref().unwrap();
        let index = self.state.index.as_ref().unwrap();
        let offset = index.offset + index.stride * first_index as wgt::BufferAddress;
        if base_vertex != 0 || first_instance != 0 {
            encoder.draw_indexed_primitives_instanced_base_instance(
                self.state.raw_primitive_type,
                index_count as _,
                index.raw_type,
                index.buffer_ptr.as_native(),
                offset,
                instance_count as _,
                base_vertex as _,
                first_instance as _,
            );
        } else if instance_count != 1 {
            encoder.draw_indexed_primitives_instanced(
                self.state.raw_primitive_type,
                index_count as _,
                index.raw_type,
                index.buffer_ptr.as_native(),
                offset,
                instance_count as _,
            );
        } else {
            encoder.draw_indexed_primitives(
                self.state.raw_primitive_type,
                index_count as _,
                index.raw_type,
                index.buffer_ptr.as_native(),
                offset,
            );
        }
    }

    unsafe fn draw_indirect(
        &mut self,
        buffer: &super::Buffer,
        mut offset: wgt::BufferAddress,
        draw_count: u32,
    ) {
        let encoder = self.state.render.as_ref().unwrap();
        for _ in 0..draw_count {
            encoder.draw_primitives_indirect(self.state.raw_primitive_type, &buffer.raw, offset);
            offset += size_of::<wgt::DrawIndirectArgs>() as wgt::BufferAddress;
        }
    }

    unsafe fn draw_indexed_indirect(
        &mut self,
        buffer: &super::Buffer,
        mut offset: wgt::BufferAddress,
        draw_count: u32,
    ) {
        let encoder = self.state.render.as_ref().unwrap();
        let index = self.state.index.as_ref().unwrap();
        for _ in 0..draw_count {
            encoder.draw_indexed_primitives_indirect(
                self.state.raw_primitive_type,
                index.raw_type,
                index.buffer_ptr.as_native(),
                index.offset,
                &buffer.raw,
                offset,
            );
            offset += size_of::<wgt::DrawIndexedIndirectArgs>() as wgt::BufferAddress;
        }
    }

    unsafe fn draw_indirect_count(
        &mut self,
        _buffer: &super::Buffer,
        _offset: wgt::BufferAddress,
        _count_buffer: &super::Buffer,
        _count_offset: wgt::BufferAddress,
        _max_count: u32,
    ) {
        //TODO
    }
    unsafe fn draw_indexed_indirect_count(
        &mut self,
        _buffer: &super::Buffer,
        _offset: wgt::BufferAddress,
        _count_buffer: &super::Buffer,
        _count_offset: wgt::BufferAddress,
        _max_count: u32,
    ) {
        //TODO
    }

    // compute

    unsafe fn begin_compute_pass(&mut self, desc: &crate::ComputePassDescriptor<super::QuerySet>) {
        self.begin_pass();

        debug_assert!(self.state.blit.is_none());
        debug_assert!(self.state.compute.is_none());
        debug_assert!(self.state.render.is_none());

        let raw = self.raw_cmd_buf.as_ref().unwrap();

        objc::rc::autoreleasepool(|| {
            // TimeStamp Queries and ComputePassDescriptor were both introduced in Metal 2.3 (macOS 11, iOS 14)
            // and we currently only need ComputePassDescriptor for timestamp queries
            let encoder = if self.shared.private_caps.timestamp_query_support.is_empty() {
                raw.new_compute_command_encoder()
            } else {
                let descriptor = metal::ComputePassDescriptor::new();

                let mut sba_index = 0;
                let mut next_sba_descriptor = || {
                    let sba_descriptor = descriptor
                        .sample_buffer_attachments()
                        .object_at(sba_index)
                        .unwrap();
                    sba_index += 1;
                    sba_descriptor
                };

                for (set, index) in self.state.pending_timer_queries.drain(..) {
                    let sba_descriptor = next_sba_descriptor();
                    sba_descriptor.set_sample_buffer(set.counter_sample_buffer.as_ref().unwrap());
                    sba_descriptor.set_start_of_encoder_sample_index(index as _);
                    sba_descriptor.set_end_of_encoder_sample_index(metal::COUNTER_DONT_SAMPLE);
                }

                if let Some(timestamp_writes) = desc.timestamp_writes.as_ref() {
                    let sba_descriptor = next_sba_descriptor();
                    sba_descriptor.set_sample_buffer(
                        timestamp_writes
                            .query_set
                            .counter_sample_buffer
                            .as_ref()
                            .unwrap(),
                    );

                    sba_descriptor.set_start_of_encoder_sample_index(
                        timestamp_writes
                            .beginning_of_pass_write_index
                            .map_or(metal::COUNTER_DONT_SAMPLE, |i| i as _),
                    );
                    sba_descriptor.set_end_of_encoder_sample_index(
                        timestamp_writes
                            .end_of_pass_write_index
                            .map_or(metal::COUNTER_DONT_SAMPLE, |i| i as _),
                    );
                }

                raw.compute_command_encoder_with_descriptor(descriptor)
            };

            if let Some(label) = desc.label {
                encoder.set_label(label);
            }

            self.state.compute = Some(encoder.to_owned());
        });
    }
    unsafe fn end_compute_pass(&mut self) {
        self.state.compute.take().unwrap().end_encoding();
    }

    unsafe fn set_compute_pipeline(&mut self, pipeline: &super::ComputePipeline) {
        self.state.raw_wg_size = pipeline.work_group_size;
        self.state.stage_infos.cs.assign_from(&pipeline.cs_info);

        let encoder = self.state.compute.as_ref().unwrap();
        encoder.set_compute_pipeline_state(&pipeline.raw);

        if let Some((index, sizes)) = self
            .state
            .make_sizes_buffer_update(naga::ShaderStage::Compute, &mut self.temp.binding_sizes)
        {
            encoder.set_bytes(
                index as _,
                (sizes.len() * WORD_SIZE) as u64,
                sizes.as_ptr().cast(),
            );
        }

        // update the threadgroup memory sizes
        while self.state.work_group_memory_sizes.len() < pipeline.work_group_memory_sizes.len() {
            self.state.work_group_memory_sizes.push(0);
        }
        for (index, (cur_size, pipeline_size)) in self
            .state
            .work_group_memory_sizes
            .iter_mut()
            .zip(pipeline.work_group_memory_sizes.iter())
            .enumerate()
        {
            const ALIGN_MASK: u32 = 0xF; // must be a multiple of 16 bytes
            let size = ((*pipeline_size - 1) | ALIGN_MASK) + 1;
            if *cur_size != size {
                *cur_size = size;
                encoder.set_threadgroup_memory_length(index as _, size as _);
            }
        }
    }

    unsafe fn dispatch(&mut self, count: [u32; 3]) {
        if count[0] > 0 && count[1] > 0 && count[2] > 0 {
            let encoder = self.state.compute.as_ref().unwrap();
            let raw_count = metal::MTLSize {
                width: count[0] as u64,
                height: count[1] as u64,
                depth: count[2] as u64,
            };
            encoder.dispatch_thread_groups(raw_count, self.state.raw_wg_size);
        }
    }

    unsafe fn dispatch_indirect(&mut self, buffer: &super::Buffer, offset: wgt::BufferAddress) {
        let encoder = self.state.compute.as_ref().unwrap();
        encoder.dispatch_thread_groups_indirect(&buffer.raw, offset, self.state.raw_wg_size);
    }

    unsafe fn build_acceleration_structures<'a, T>(
        &mut self,
        _descriptor_count: u32,
        _descriptors: T,
    ) where
        super::Api: 'a,
        T: IntoIterator<
            Item = crate::BuildAccelerationStructureDescriptor<
                'a,
                super::Buffer,
                super::AccelerationStructure,
            >,
        >,
    {
        unimplemented!()
    }

    unsafe fn place_acceleration_structure_barrier(
        &mut self,
        _barriers: crate::AccelerationStructureBarrier,
    ) {
        unimplemented!()
    }
}

impl Drop for super::CommandEncoder {
    fn drop(&mut self) {
        // Metal raises an assert when a MTLCommandEncoder is deallocated without a call
        // to endEncoding. This isn't documented in the general case at
        // https://developer.apple.com/documentation/metal/mtlcommandencoder, but for the
        // more-specific MTLComputeCommandEncoder it is stated as a requirement at
        // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder. It
        // appears to be a requirement for all MTLCommandEncoder objects. Failing to call
        // endEncoding causes a crash with the message 'Command encoder released without
        // endEncoding'. To prevent this, we explicitiy call discard_encoding, which
        // calls end_encoding on any still-held metal::CommandEncoders.
        unsafe {
            self.discard_encoding();
        }
        self.counters.command_encoders.sub(1);
    }
}